doccupine 0.0.119 → 0.0.121

package/README.md

@@ -45,11 +45,12 @@ doccupine config --reset    # Re-prompt for configuration
 
 `watch` (default):
 
-| Flag            | Description                                                          |
-| --------------- | -------------------------------------------------------------------- |
-| `--port <port>` | Port for the dev server (default: `3000`). Auto-increments if taken. |
-| `--verbose`     | Show all Next.js output including compilation details                |
-| `--reset`       | Re-prompt for watch/output directories                               |
+| Flag                       | Description                                                                                                                              |
+| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
+| `--port <port>`            | Port for the dev server (default: `3000`). Auto-increments if taken.                                                                     |
+| `--verbose`                | Show all Next.js output including compilation details                                                                                    |
+| `--reset`                  | Re-prompt for watch/output directories                                                                                                   |
+| `--package-manager <name>` | Package manager for the generated app: `pnpm` or `npm` (default: auto-detect). Overrides the `packageManager` field in `doccupine.json`. |
 
 `build`:
 
@@ -75,10 +76,13 @@ description: "Page description for SEO"
 category: "Getting Started"
 categoryOrder: 0 # Sort order for the category group
 order: 1 # Sort order within the category
+navIcon: "book-open" # Lucide icon name shown next to the page in the sidebar
+categoryIcon: "rocket" # Lucide icon name shown next to the category group
 name: "My Docs" # Override site name in title suffix
 icon: "https://..." # Page favicon URL
 image: "https://..." # OpenGraph image URL
 date: "2025-01-01" # Page date metadata
+updated: "2025-02-01" # Last-modified date (JSON-LD dateModified)
 section: "API Reference" # Section this page belongs to
 sectionOrder: 1 # Sort order for the section in the tab bar
 ---

package/dist/lib/structures.js

@@ -58,6 +58,7 @@ import { mcpIndexTemplate } from "../templates/services/mcp/index.js";
 import { mcpServerTemplate } from "../templates/services/mcp/server.js";
 import { mcpToolsTemplate } from "../templates/services/mcp/tools.js";
 import { mcpTypesTemplate } from "../templates/services/mcp/types.js";
+import { vectorHelpersTemplate } from "../templates/services/mcp/vector.js";
 import { docsIndexStubTemplate } from "../templates/services/mcp/docsIndexStub.js";
 import { llmConfigTemplate } from "../templates/services/llm/config.js";
 import { llmFactoryTemplate } from "../templates/services/llm/factory.js";
@@ -138,6 +139,7 @@ export const appStructure = {
     "services/mcp/server.ts": mcpServerTemplate,
     "services/mcp/tools.ts": mcpToolsTemplate,
     "services/mcp/types.ts": mcpTypesTemplate,
+    "services/mcp/vector.ts": vectorHelpersTemplate,
     "services/mcp/docs-index.json": docsIndexStubTemplate,
     "services/llm/config.ts": llmConfigTemplate,
     "services/llm/factory.ts": llmFactoryTemplate,

package/dist/templates/app/api/rag/route.d.ts

@@ -1 +1 @@
-export declare const ragRoutesTemplate = "import { NextResponse } from \"next/server\";\nimport { z } from \"zod\";\nimport { getLLMConfig, createChatModel } from \"@/services/llm\";\nimport {\n  searchDocs,\n  ensureDocsIndex,\n  getIndexStatus,\n} from \"@/services/mcp/server\";\nimport { rateLimit } from \"@/utils/rateLimit\";\nimport { config } from \"@/utils/config\";\n\nconst messageSchema = z.object({\n  role: z.enum([\"user\", \"assistant\"]),\n  content: z.string().max(4000),\n});\n\nconst ragSchema = z.object({\n  question: z.string().min(1).max(2000),\n  history: z.array(messageSchema).max(20).optional(),\n  refresh: z.boolean().optional(),\n});\n\nconst projectName = config.name || \"Doccupine\";\n\nconst systemContext = `You are AI Assistant, a documentation assistant for ${projectName}, Your name is ${projectName} AI Assistant.\n\n## Core Rules\n1. Answer ONLY from the provided context. Never fabricate information.\n2. If the answer isn't in the context, say so clearly and suggest relevant sections or pages the user might check.\n3. If the question is ambiguous, ask a brief clarifying question before answering.\n\n## Response Style\n- Be concise and direct. Lead with the answer, then provide details if needed.\n- Use code examples from the context when relevant.\n- Match the technical level of the user's question.\n\n## MDX/Code Formatting\nWhen including code blocks in your response:\n- Never nest fenced code blocks (triple backticks) inside other fenced code blocks.\n- If you need to show MDX source that itself contains code blocks, use indented code blocks or escape the inner backticks.\n- All output must be valid MDX that renders correctly.\n\n## Internal Links\nEach context chunk includes a \"URL:\" line with the pre-computed page URL. Use it directly when linking:\n- Format links as markdown: [Page Title](/slug/).\n- Never expose raw file paths like \"/app/.../page.tsx\" to the user.\n- Never include route group segments in parentheses, like \"(site)\", in any link - they are internal folder names and do not exist in real URLs. Use \"/code/\" not \"/(site)/code/\".\n- Do NOT add a \"Related Pages\" section at the end - sources are shown separately by the UI.\n\n## Greetings & Small Talk\nIf the user sends a greeting or non-documentation question, respond briefly and ask how you can help with the documentation.`;\n\n// LangChain + the MCP SDK require the Node.js runtime (not edge).\nexport const runtime = \"nodejs\";\n// Safety net for the streaming function (Vercel-only; ignored elsewhere).\n// The heartbeat below, not this value, is what prevents proxy 524 timeouts.\nexport const maxDuration = 60;\n\nexport async function POST(req: Request) {\n  // Rate limit by IP\n  const ip =\n    req.headers.get(\"x-forwarded-for\")?.split(\",\")[0]?.trim() ?? \"unknown\";\n  const { allowed, retryAfter } = rateLimit(ip);\n  if (!allowed) {\n    return NextResponse.json(\n      { error: \"Too many requests\" },\n      { status: 429, headers: { \"Retry-After\": String(retryAfter) } },\n    );\n  }\n\n  // Validate the request up front so genuine client/config errors still return\n  // real status codes before we commit to a 200 streaming response.\n  let body: unknown;\n  try {\n    body = await req.json();\n  } catch {\n    return NextResponse.json({ error: \"Invalid JSON body\" }, { status: 400 });\n  }\n\n  const parsed = ragSchema.safeParse(body);\n  if (!parsed.success) {\n    return NextResponse.json(\n      { error: \"Invalid input\", details: parsed.error.issues },\n      { status: 400 },\n    );\n  }\n  const { question, history, refresh } = parsed.data;\n\n  let llmConfig;\n  try {\n    llmConfig = getLLMConfig();\n  } catch (error: unknown) {\n    const message =\n      error instanceof Error ? error.message : \"LLM configuration error\";\n    return NextResponse.json({ error: message }, { status: 500 });\n  }\n\n  const encoder = new TextEncoder();\n  let heartbeat: ReturnType<typeof setInterval> | null = null;\n  let streamClosed = false;\n\n  // Return the streaming response immediately and do ALL slow work (indexing,\n  // search, model streaming) inside start(). This flushes headers + a first\n  // byte right away so edge proxies never hit their time-to-first-byte timeout.\n  const readableStream = new ReadableStream({\n    async start(controller) {\n      const safeEnqueue = (payload: string) => {\n        if (streamClosed) return;\n        try {\n          controller.enqueue(encoder.encode(payload));\n        } catch {\n          // Stream already closed or cancelled - stop emitting.\n          streamClosed = true;\n          if (heartbeat) clearInterval(heartbeat);\n        }\n      };\n\n      // Keep the connection alive while the index and model warm up.\n      heartbeat = setInterval(() => safeEnqueue(`: keep-alive\\n\\n`), 15000);\n\n      try {\n        // First byte, before any slow work - satisfies the proxy TTFB window.\n        safeEnqueue(`: connected\\n\\n`);\n\n        // Ensure docs are indexed (loads precomputed embeddings when present).\n        await ensureDocsIndex(Boolean(refresh));\n\n        // Use MCP search_docs tool to find relevant documentation\n        const searchResults = await searchDocs(question, 6);\n\n        // Build context from search results\n        const context = searchResults\n          .map(({ chunk, score }) => {\n            const slug = chunk.uri.replace(\"docs://\", \"\").replace(/^\\/+/, \"\");\n            const url = slug ? `/${slug}/` : \"/\";\n            return `File: ${chunk.path}\\nURL: ${url}\\nScore: ${score.toFixed(3)}\\n----\\n${chunk.text}`;\n          })\n          .join(\"\\n\\n================\\n\\n\");\n\n        // Build metadata from MCP search results and send it before the answer.\n        const indexStatus = getIndexStatus();\n        const metadata = {\n          sources: searchResults.map(({ chunk, score }) => ({\n            id: chunk.id,\n            path: chunk.path,\n            uri: chunk.uri,\n            score,\n          })),\n          chunkCount: indexStatus.chunkCount,\n        };\n        safeEnqueue(\n          `data: ${JSON.stringify({ type: \"metadata\", data: metadata })}\\n\\n`,\n        );\n\n        // Assemble the prompt, including conversation history for multi-turn.\n        const prompt: {\n          role: \"system\" | \"user\" | \"assistant\";\n          content: string;\n        }[] = [\n          {\n            role: \"system\" as const,\n            content: systemContext,\n          },\n        ];\n        if (history && history.length > 0) {\n          for (const msg of history) {\n            prompt.push({\n              role: msg.role,\n              content: msg.content,\n            });\n          }\n        }\n        prompt.push({\n          role: \"user\" as const,\n          content: `Question: ${question}\\n\\nContext:\\n${context}`,\n        });\n\n        // Create chat model and stream the response.\n        const llm = createChatModel(llmConfig);\n        const stream = await llm.stream(prompt);\n        for await (const chunk of stream) {\n          const content = chunk?.content || \"\";\n          if (content) {\n            safeEnqueue(\n              `data: ${JSON.stringify({ type: \"content\", data: content })}\\n\\n`,\n            );\n          }\n        }\n\n        safeEnqueue(`data: ${JSON.stringify({ type: \"done\" })}\\n\\n`);\n      } catch (error: unknown) {\n        const message = error instanceof Error ? error.message : \"Stream error\";\n        safeEnqueue(\n          `data: ${JSON.stringify({ type: \"error\", data: message })}\\n\\n`,\n        );\n      } finally {\n        if (heartbeat) clearInterval(heartbeat);\n        streamClosed = true;\n        try {\n          controller.close();\n        } catch {\n          // Already closed.\n        }\n      }\n    },\n    cancel() {\n      streamClosed = true;\n      if (heartbeat) clearInterval(heartbeat);\n    },\n  });\n\n  return new Response(readableStream, {\n    headers: {\n      \"Content-Type\": \"text/event-stream\",\n      \"Cache-Control\": \"no-cache, no-transform\",\n    },\n  });\n}\n\nexport async function GET() {\n  const status = getIndexStatus();\n  return NextResponse.json({\n    ready: status.ready,\n    chunks: status.chunkCount,\n  });\n}\n";
+export declare const ragRoutesTemplate = "import { NextResponse } from \"next/server\";\nimport { z } from \"zod\";\nimport { getLLMConfig, createChatModel } from \"@/services/llm\";\nimport {\n  searchDocs,\n  ensureDocsIndex,\n  getIndexStatus,\n} from \"@/services/mcp/server\";\nimport { rateLimit } from \"@/utils/rateLimit\";\nimport { config } from \"@/utils/config\";\n\nconst messageSchema = z.object({\n  role: z.enum([\"user\", \"assistant\"]),\n  content: z.string().max(4000),\n});\n\nconst ragSchema = z.object({\n  question: z.string().min(1).max(2000),\n  history: z.array(messageSchema).max(20).optional(),\n  refresh: z.boolean().optional(),\n});\n\nconst projectName = config.name || \"Doccupine\";\n\nconst systemContext = `You are AI Assistant, a documentation assistant for ${projectName}, Your name is ${projectName} AI Assistant.\n\n## Core Rules\n1. Answer ONLY from the provided context. Never fabricate information.\n2. If the answer isn't in the context, say so clearly and suggest relevant sections or pages the user might check.\n3. If the question is ambiguous, ask a brief clarifying question before answering.\n\n## Response Style\n- Be concise and direct. Lead with the answer, then provide details if needed.\n- Use code examples from the context when relevant.\n- Match the technical level of the user's question.\n\n## MDX/Code Formatting\nWhen including code blocks in your response:\n- Never nest fenced code blocks (triple backticks) inside other fenced code blocks.\n- If you need to show MDX source that itself contains code blocks, use indented code blocks or escape the inner backticks.\n- All output must be valid MDX that renders correctly.\n\n## Internal Links\nEach context chunk includes a \"URL:\" line with the pre-computed page URL. Use it directly when linking:\n- Format links as markdown: [Page Title](/slug/).\n- Never expose raw file paths like \"/app/.../page.tsx\" to the user.\n- Never include route group segments in parentheses, like \"(site)\", in any link - they are internal folder names and do not exist in real URLs. Use \"/code/\" not \"/(site)/code/\".\n- Do NOT add a \"Related Pages\" section at the end - sources are shown separately by the UI.\n\n## Greetings & Small Talk\nIf the user sends a greeting or non-documentation question, respond briefly and ask how you can help with the documentation.`;\n\n// LangChain + the MCP SDK require the Node.js runtime (not edge).\nexport const runtime = \"nodejs\";\n// Safety net for the streaming function (Vercel-only; ignored elsewhere).\n// The heartbeat below, not this value, is what prevents proxy 524 timeouts.\nexport const maxDuration = 60;\n\nexport async function POST(req: Request) {\n  // Rate limit by IP\n  const ip =\n    req.headers.get(\"x-forwarded-for\")?.split(\",\")[0]?.trim() ?? \"unknown\";\n  const { allowed, retryAfter } = rateLimit(ip);\n  if (!allowed) {\n    return NextResponse.json(\n      { error: \"Too many requests\" },\n      { status: 429, headers: { \"Retry-After\": String(retryAfter) } },\n    );\n  }\n\n  // Validate the request up front so genuine client/config errors still return\n  // real status codes before we commit to a 200 streaming response.\n  let body: unknown;\n  try {\n    body = await req.json();\n  } catch {\n    return NextResponse.json({ error: \"Invalid JSON body\" }, { status: 400 });\n  }\n\n  const parsed = ragSchema.safeParse(body);\n  if (!parsed.success) {\n    return NextResponse.json(\n      { error: \"Invalid input\", details: parsed.error.issues },\n      { status: 400 },\n    );\n  }\n  const { question, history, refresh } = parsed.data;\n\n  let llmConfig;\n  try {\n    llmConfig = getLLMConfig();\n  } catch (error: unknown) {\n    const message =\n      error instanceof Error ? error.message : \"LLM configuration error\";\n    return NextResponse.json({ error: message }, { status: 500 });\n  }\n\n  const encoder = new TextEncoder();\n  let heartbeat: ReturnType<typeof setInterval> | null = null;\n  let streamClosed = false;\n\n  // Return the streaming response immediately and do ALL slow work (indexing,\n  // search, model streaming) inside start(). This flushes headers + a first\n  // byte right away so edge proxies never hit their time-to-first-byte timeout.\n  const readableStream = new ReadableStream({\n    async start(controller) {\n      const safeEnqueue = (payload: string) => {\n        if (streamClosed) return;\n        try {\n          controller.enqueue(encoder.encode(payload));\n        } catch {\n          // Stream already closed or cancelled - stop emitting.\n          streamClosed = true;\n          if (heartbeat) clearInterval(heartbeat);\n        }\n      };\n\n      // Keep the connection alive while the index and model warm up.\n      heartbeat = setInterval(() => safeEnqueue(`: keep-alive\\n\\n`), 15000);\n\n      try {\n        // First byte, before any slow work - satisfies the proxy TTFB window.\n        safeEnqueue(`: connected\\n\\n`);\n\n        // Ensure docs are indexed (loads precomputed embeddings when present).\n        await ensureDocsIndex(Boolean(refresh));\n\n        // Use MCP search_docs tool to find relevant documentation\n        const searchResults = await searchDocs(question, 6);\n\n        // Build context from search results\n        const context = searchResults\n          .map(({ chunk, score }) => {\n            const slug = chunk.uri.replace(\"docs://\", \"\").replace(/^\\/+/, \"\");\n            const url = slug ? `/${slug}/` : \"/\";\n            return `File: ${chunk.path}\\nURL: ${url}\\nScore: ${score.toFixed(3)}\\n----\\n${chunk.text}`;\n          })\n          .join(\"\\n\\n================\\n\\n\");\n\n        // Build metadata from MCP search results and send it before the answer.\n        const indexStatus = getIndexStatus();\n        const metadata = {\n          sources: searchResults.map(({ chunk, score }) => ({\n            id: chunk.id,\n            path: chunk.path,\n            uri: chunk.uri,\n            score,\n          })),\n          chunkCount: indexStatus.chunkCount,\n        };\n        safeEnqueue(\n          `data: ${JSON.stringify({ type: \"metadata\", data: metadata })}\\n\\n`,\n        );\n\n        // Assemble the prompt, including conversation history for multi-turn.\n        const prompt: {\n          role: \"system\" | \"user\" | \"assistant\";\n          content: string;\n        }[] = [\n          {\n            role: \"system\" as const,\n            content: systemContext,\n          },\n        ];\n        if (history && history.length > 0) {\n          for (const msg of history) {\n            prompt.push({\n              role: msg.role,\n              content: msg.content,\n            });\n          }\n        }\n        prompt.push({\n          role: \"user\" as const,\n          content: `Question: ${question}\\n\\nContext:\\n${context}`,\n        });\n\n        // Create chat model and stream the response.\n        const llm = createChatModel(llmConfig);\n        const stream = await llm.stream(prompt);\n        for await (const chunk of stream) {\n          const content = chunk?.content || \"\";\n          if (content) {\n            safeEnqueue(\n              `data: ${JSON.stringify({ type: \"content\", data: content })}\\n\\n`,\n            );\n          }\n        }\n\n        safeEnqueue(`data: ${JSON.stringify({ type: \"done\" })}\\n\\n`);\n      } catch (error: unknown) {\n        const message = error instanceof Error ? error.message : \"Stream error\";\n        safeEnqueue(\n          `data: ${JSON.stringify({ type: \"error\", data: message })}\\n\\n`,\n        );\n      } finally {\n        if (heartbeat) clearInterval(heartbeat);\n        streamClosed = true;\n        try {\n          controller.close();\n        } catch {\n          // Already closed.\n        }\n      }\n    },\n    cancel() {\n      streamClosed = true;\n      if (heartbeat) clearInterval(heartbeat);\n    },\n  });\n\n  return new Response(readableStream, {\n    headers: {\n      \"Content-Type\": \"text/event-stream\",\n      \"Cache-Control\": \"no-cache, no-transform\",\n    },\n  });\n}\n\nexport async function GET() {\n  const status = getIndexStatus();\n  return NextResponse.json({\n    ready: status.ready,\n    chunks: status.chunkCount,\n    reason: status.reason,\n  });\n}\n";

package/dist/templates/app/api/rag/route.js

@@ -222,6 +222,7 @@ export async function GET() {
   return NextResponse.json({
     ready: status.ready,
     chunks: status.chunkCount,
+    reason: status.reason,
   });
 }
 `;

package/dist/templates/env.example.d.ts

@@ -1 +1 @@
-export declare const envExampleTemplate = "# Public Site URL\n# Used by sitemap.xml and robots.txt. Overrides `url` in config.json when set.\n# NEXT_PUBLIC_SITE_URL=https://docs.example.com\n\n# Password Protection (optional)\n# Set a shared password to gate the whole site behind a login screen. When set,\n# pages require the password, the content APIs (chat + search) return 401\n# without it, and the site is hidden from search engines and crawlers. Leave\n# unset (or remove) to keep the site public.\n# SITE_PASSWORD=choose-a-strong-shared-password\n\n# LLM Provider Configuration\n# Choose your preferred LLM provider: openai, anthropic, or google\nLLM_PROVIDER=openai\n\n# API Keys (set the one matching your provider)\nOPENAI_API_KEY=your_openai_api_key_here\nANTHROPIC_API_KEY=your_anthropic_api_key_here\nGOOGLE_API_KEY=your_google_api_key_here\n\n# Optional: Override default chat model\n# See available models at your provider's docs:\n#   OpenAI:    https://platform.openai.com/docs/models\n#   Anthropic: https://docs.anthropic.com/claude/docs/models-overview\n#   Google:    https://ai.google.dev/models/gemini\n# LLM_CHAT_MODEL=\n\n# Optional: Override default embedding model\n# See available embedding models at your provider's docs:\n#   OpenAI: https://platform.openai.com/docs/guides/embeddings\n#   Google: https://ai.google.dev/gemini-api/docs/embeddings\n# Note: Anthropic doesn't provide embeddings, will fallback to OpenAI\n# LLM_EMBEDDING_MODEL=\n\n# Optional: Set temperature (0-1, default: 0)\n# LLM_TEMPERATURE=0\n";
+export declare const envExampleTemplate = "# Public Site URL\n# Used by sitemap.xml and robots.txt. Overrides `url` in config.json when set.\n# NEXT_PUBLIC_SITE_URL=https://docs.example.com\n\n# Password Protection (optional)\n# Set a shared password to gate the whole site behind a login screen. When set,\n# pages require the password, the content APIs (chat + search) return 401\n# without it, and the site is hidden from search engines and crawlers. Leave\n# unset (or remove) to keep the site public.\n# SITE_PASSWORD=choose-a-strong-shared-password\n\n# LLM Provider Configuration\n# Choose your preferred LLM provider: openai, anthropic, or google\nLLM_PROVIDER=openai\n\n# API Keys (set the one matching your provider)\nOPENAI_API_KEY=your_openai_api_key_here\nANTHROPIC_API_KEY=your_anthropic_api_key_here\nGOOGLE_API_KEY=your_google_api_key_here\n\n# Optional: Override default chat model\n# See available models at your provider's docs:\n#   OpenAI:    https://platform.openai.com/docs/models\n#   Anthropic: https://docs.anthropic.com/claude/docs/models-overview\n#   Google:    https://ai.google.dev/models/gemini\n# LLM_CHAT_MODEL=\n\n# Optional: Override default embedding model\n# See available embedding models at your provider's docs:\n#   OpenAI: https://platform.openai.com/docs/guides/embeddings\n#   Google: https://ai.google.dev/gemini-api/docs/embeddings\n# Note: Anthropic doesn't provide embeddings, will fallback to OpenAI\n# LLM_EMBEDDING_MODEL=\n\n# Optional: Set temperature (0-1, default: 0)\n# LLM_TEMPERATURE=0\n\n# Optional: Embedding dimensions for the prebuilt search index (default: 512)\n# Doc vectors are Matryoshka-truncated to this many dimensions and stored as\n# int8, which keeps services/mcp/docs-index.json small (~20x smaller than raw\n# floats) so large doc sets don't stall the AI chat on serverless cold starts.\n# Lower = smaller index, slightly lower recall. Values >= the model's native\n# dimension keep full precision. Rebuild after changing this.\n# LLM_EMBEDDING_DIMS=512\n";

package/dist/templates/env.example.js

@@ -34,4 +34,12 @@ GOOGLE_API_KEY=your_google_api_key_here
 
 # Optional: Set temperature (0-1, default: 0)
 # LLM_TEMPERATURE=0
+
+# Optional: Embedding dimensions for the prebuilt search index (default: 512)
+# Doc vectors are Matryoshka-truncated to this many dimensions and stored as
+# int8, which keeps services/mcp/docs-index.json small (~20x smaller than raw
+# floats) so large doc sets don't stall the AI chat on serverless cold starts.
+# Lower = smaller index, slightly lower recall. Values >= the model's native
+# dimension keep full precision. Rebuild after changing this.
+# LLM_EMBEDDING_DIMS=512
 `;

package/dist/templates/mdx/model-context-protocol.mdx.d.ts

@@ -1 +1 @@
-export declare const mcpMdxTemplate = "---\ntitle: \"Model Context Protocol\"\ndescription: \"Connect your Doccupine documentation to AI tools with an MCP server for enhanced AI-powered documentation search.\"\ndate: \"2026-02-19\"\ncategory: \"Configuration\"\ncategoryOrder: 3\norder: 9\n---\n\n# Model Context Protocol\n\nConnect your documentation to AI tools with a hosted MCP server.\n\nDoccupine automatically generates a Model Context Protocol (MCP) server from your documentation, making your content accessible to AI applications like Claude, Cursor, VS Code, and other MCP-compatible tools. Your MCP server exposes semantic search capabilities, allowing AI tools to query your documentation directly and provide accurate, context-aware answers.\n\n<Callout type=\"warning\">\n  The MCP server requires the [AI Assistant](/ai-assistant) to be configured. Make sure you have set up your LLM provider and API keys before using the MCP server.\n</Callout>\n\n## About MCP servers\n\nThe Model Context Protocol (MCP) is an open protocol that creates standardized connections between AI applications and external services, like documentation. Doccupine generates an MCP server from your documentation, preparing your content for the broader AI ecosystem where any MCP client can connect to your documentation.\n\nYour MCP server exposes search and retrieval tools for AI applications to query your documentation. Your users must connect your MCP server to their preferred AI tools to access your documentation.\n\n### How MCP servers work\n\nWhen an AI tool has your documentation MCP server connected, the AI tool can search your documentation directly instead of making a generic web search in response to a user's prompt. Your MCP server provides access to all indexed content from your documentation site.\n\n- The LLM can proactively search your documentation while generating a response, not just when explicitly asked.\n- The LLM determines when to use the search tool based on the context of the conversation and the relevance of your documentation.\n- Each tool call happens during the generation process, so the LLM searches up-to-date information from your documentation to generate its response.\n\n### MCP compared to web search\n\nAI tools can search the web, but MCP provides distinct advantages for documentation.\n\n- **Direct source access**: Web search depends on what search engines have indexed, which may be stale or incomplete. MCP searches your current indexed documentation directly.\n- **Integrated workflow**: MCP allows the AI to search during response generation rather than performing a separate web search.\n- **Semantic search**: MCP uses vector embeddings for semantic similarity search, providing more relevant results than keyword-based web search.\n- **No search noise**: SEO and ranking algorithms influence web search results. MCP goes straight to your documentation content.\n\n## Access your MCP server\n\nDoccupine automatically generates an MCP server for your documentation and hosts it at your documentation URL with the `/api/mcp` path. For example, if your documentation is hosted at `https://example.com`, your MCP server is available at `https://example.com/api/mcp`.\n\nThe MCP server provides both a GET endpoint to discover available tools and a POST endpoint to execute tool calls.\n\n### Authentication\n\nYou can optionally protect your MCP server with an API key by setting the `DOCS_API_KEY` environment variable in your `.env` file:\n\n```bash\nDOCS_API_KEY=your-secret-key\n```\n\nWhen `DOCS_API_KEY` is set, all requests to `/api/mcp` must include an `Authorization` header with a Bearer token:\n\n```text\nAuthorization: Bearer your-secret-key\n```\n\nRequests without a valid token receive a `401 Unauthorized` response. When `DOCS_API_KEY` is not set, the MCP server is publicly accessible with no authentication required.\n\n<Callout type=\"note\">\n  Authentication only applies to the `/api/mcp` endpoint. The `/api/rag` endpoint used by the built-in AI Assistant chat is not affected by this setting.\n</Callout>\n\n### API Endpoints\n\n#### GET /api/mcp\n\nReturns information about available tools and the current index status.\n\n**Response:**\n\n```json\n{\n  \"tools\": [\n    {\n      \"name\": \"search_docs\",\n      \"description\": \"Search through the documentation content using semantic search...\",\n      \"inputSchema\": { ... }\n    },\n    ...\n  ],\n  \"index\": {\n    \"ready\": true,\n    \"chunkCount\": 150\n  }\n}\n```\n\n#### POST /api/mcp\n\nExecutes an MCP tool call.\n\n**Request Body:**\n\n```json\n{\n  \"tool\": \"search_docs\",\n  \"params\": {\n    \"query\": \"how to deploy\",\n    \"limit\": 6\n  }\n}\n```\n\n**Response:**\n\n```json\n{\n  \"content\": [\n    {\n      \"path\": \"app/deployment-and-hosting/page.tsx\",\n      \"uri\": \"docs://deployment-and-hosting\",\n      \"score\": \"0.892\",\n      \"text\": \"Deploy your Doccupine site as a Next.js application...\"\n    },\n    ...\n  ]\n}\n```\n\n## Available Tools\n\nYour MCP server exposes three tools for interacting with your documentation:\n\n### search_docs\n\nSearch through the documentation content using semantic search. Returns relevant chunks of documentation based on the query using vector embeddings and cosine similarity.\n\n**Parameters:**\n\n- `query` (required): The search query to find relevant documentation\n- `limit` (optional): Maximum number of results to return (default: 6)\n\n**Example:**\n\n```json\n{\n  \"tool\": \"search_docs\",\n  \"params\": {\n    \"query\": \"how to configure AI assistant\",\n    \"limit\": 5\n  }\n}\n```\n\n### get_doc\n\nGet the full content of a specific documentation page by its path.\n\n**Parameters:**\n\n- `path` (required): The file path to the documentation page (e.g., `app/getting-started/page.tsx`)\n\n**Example:**\n\n```json\n{\n  \"tool\": \"get_doc\",\n  \"params\": {\n    \"path\": \"app/deployment-and-hosting/page.tsx\"\n  }\n}\n```\n\n### list_docs\n\nList all available documentation pages, optionally filtered by directory.\n\n**Parameters:**\n\n- `directory` (optional): Optional directory to filter results (e.g., `components`)\n\n**Example:**\n\n```json\n{\n  \"tool\": \"list_docs\",\n  \"params\": {\n    \"directory\": \"configuration\"\n  }\n}\n```\n\n## How it works\n\nDoccupine's MCP server uses semantic search powered by vector embeddings to provide accurate, context-aware search results.\n\n### Indexing Process\n\nDoccupine runs this pipeline at build time and ships the resulting vectors with your site, so the server loads a ready-made index instead of re-embedding your docs on every cold start.\n\n1. **Document Discovery**: Doccupine scans your `app/` directory for all `page.tsx`, `page.ts`, `page.jsx`, and `page.js` files.\n2. **Content Extraction**: It extracts content from `const content =` declarations in your page files.\n3. **Chunking**: Large documents are split into chunks of approximately 800 characters with 100 character overlap for better context preservation.\n4. **Embedding Generation**: Each chunk is converted to a vector embedding using your configured LLM provider's embedding model.\n5. **Index Building**: The embeddings are written to `services/mcp/docs-index.json`, bundled into your serverless functions, and loaded into memory at runtime for fast similarity search.\n\n### Search Process\n\n1. **Query Embedding**: The search query is converted to a vector embedding using the same embedding model.\n2. **Similarity Calculation**: Cosine similarity is calculated between the query embedding and all document chunk embeddings.\n3. **Ranking**: Results are sorted by similarity score (highest first).\n4. **Response**: The top N results (based on the limit parameter) are returned with their paths, URIs, scores, and text content.\n\n<Callout type=\"note\">\n  Embeddings are precomputed at build time and bundled with your site, so the server loads the ready-made index into memory on startup instead of re-embedding your docs on every cold start. It only embeds at runtime when the precomputed index is missing or was built with a different provider or model. When you update your documentation, rebuild or redeploy your site to regenerate the index with fresh content.\n</Callout>\n\n## Use your MCP server\n\nYour users must connect your MCP server to their preferred AI tools.\n\n1. Make your MCP server URL publicly available.\n2. Users copy your MCP server URL and add it to their tools.\n3. Users access your documentation through their tools.\n\nThese are some of the ways you can help your users connect to your MCP server:\n\n<Tabs>\n  <TabContent title=\"Claude\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the steps to connect it to Claude.\n        1. Navigate to the [Connectors](https://claude.ai/settings/connectors) page in the Claude settings.\n        2. Select **Add custom connector**.\n        3. Add your MCP server name and URL.\n        4. Select **Add**.\n        5. When using Claude, select the attachments button (the plus icon).\n        6. Select your MCP server.\n      </Step>\n    </Steps>\n    See the [Model Context Protocol documentation](https://modelcontextprotocol.io/docs/tutorials/use-remote-mcp-server#connecting-to-a-remote-mcp-server) for more details.\n  </TabContent>\n  <TabContent title=\"Claude Code\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the command to connect it to Claude Code.\n        ```bash\n        claude mcp add --transport http <name> <url>\n        ```\n      </Step>\n    </Steps>\n    See the [Claude Code documentation](https://docs.anthropic.com/en/docs/claude-code/mcp#installing-mcp-servers) for more details.\n  </TabContent>\n  <TabContent title=\"Cursor\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the steps to connect it to Cursor.\n        1. Use <kbd>Command</kbd> + <kbd>Shift</kbd> + <kbd>P</kbd> (<kbd>Ctrl</kbd> + <kbd>Shift</kbd> + <kbd>P</kbd> on Windows) to open the command palette.\n        2. Search for \"Open MCP settings\".\n        3. Select **Add custom MCP**. This opens the `mcp.json` file.\n        4. In `mcp.json`, configure your server:\n        ```json\n        {\n          \"mcpServers\": {\n            \"<your-mcp-server-name>\": {\n              \"url\": \"<your-mcp-server-url>\"\n            }\n          }\n        }\n        ```\n      </Step>\n    </Steps>\n    See the [Cursor documentation](https://docs.cursor.com/en/context/mcp#installing-mcp-servers) for more details.\n  </TabContent>\n  <TabContent title=\"VS Code\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the steps to connect it to VS Code.\n        1. Create a `.vscode/mcp.json` file.\n        2. In `mcp.json`, configure your server:\n        ```json\n        {\n          \"servers\": {\n            \"<your-mcp-server-name>\": {\n              \"type\": \"http\",\n              \"url\": \"<your-mcp-server-url>\"\n            }\n          }\n        }\n        ```\n      </Step>\n    </Steps>\n    See the [VS Code documentation](https://code.visualstudio.com/docs/copilot/chat/mcp-servers) for more details.\n  </TabContent>\n</Tabs>\n\n## Requirements\n\nTo use the MCP server, you need to have the AI Assistant configured. The MCP server uses the same LLM configuration for generating embeddings.\n\n| Variable       | Required | Description                                                  |\n| -------------- | -------- | ------------------------------------------------------------ |\n| `LLM_PROVIDER` | Yes      | Your LLM provider (`openai`, `anthropic`, or `google`)       |\n| `DOCS_API_KEY` | No       | When set, requires Bearer token authentication on `/api/mcp` |\n\n<Callout type=\"warning\">\n  The MCP server requires an LLM provider to be configured for generating embeddings. Make sure you have set up your AI Assistant with a valid API key before using the MCP server.\n</Callout>\n\nSee the [AI Assistant documentation](/ai-assistant) for configuration details.\n\n## Content indexing\n\nYour MCP server searches content extracted from your page files. The server automatically discovers and indexes all `page.tsx`, `page.ts`, `page.jsx`, and `page.js` files in your `app/` directory.\n\n### Content extraction\n\nThe server extracts content from `const content =` declarations in your page files. Make sure your documentation pages export a `content` constant with your markdown or MDX content.\n\n**Example:**\n\n```tsx\nexport const content = `\n# Getting Started\n\nWelcome to the documentation...\n`;\n```\n\n### Excluded directories\n\nThe following directories are automatically excluded from indexing:\n\n- `node_modules`\n- `.next`\n- `.git`\n- `api`\n\n## Troubleshooting\n\n### Index not building\n\nIf the index is not building, check:\n\n- Your LLM provider is configured correctly in your `.env` file\n- You have a valid API key set\n- Your documentation pages export a `content` constant\n\n### No search results\n\nIf searches return no results:\n\n- Verify that your documentation pages are in the `app/` directory\n- Check that your pages export a `content` constant\n- Ensure the index has been built (check the `index.ready` status via GET `/api/mcp`)\n\n### Slow search performance\n\nSearches use the precomputed in-memory index and are fast. The first search is only slow when the app has to embed your docs at runtime, which happens when the precomputed index is missing or was built with a different provider or model. If performance is consistently slow:\n\n- Confirm the build ran the doc-index precompute step (the `build` script runs it before `next build`) with a valid API key\n- Check your embedding API response times\n- Consider reducing the number of documentation pages\n- Verify your server has sufficient memory\n\n### Cloudflare blocking MCP requests\n\nIf you use Cloudflare as a proxy in front of your documentation site, Cloudflare's bot protection may block server-to-server requests from AI tools like Claude.ai. This can cause MCP connections to fail silently or return errors.\n\nThere are two ways to resolve this:\n\n**Option 1: Disable the Cloudflare proxy (simplest)**\n\nIn your Cloudflare DNS settings, click the orange cloud icon next to your domain record to switch it to \"DNS only\" (grey cloud). This disables Cloudflare's proxy and bot protection for your domain, allowing MCP requests to reach your server directly.\n\n**Option 2: Add a Cloudflare WAF exception (keeps your custom domain proxied)**\n\nIn Cloudflare dashboard:\n\n1. Go to **Security > WAF**.\n2. Click **Create rule**.\n3. Set it up as:\n   - **Rule name**: Allow MCP API\n   - **Field**: URI Path\n   - **Operator**: starts with\n   - **Value**: `/api/mcp`\n   - **Action**: Skip -- then check all remaining custom rules, Rate limiting rules, and Bot Fight Mode / Super Bot Fight Mode.\n4. Deploy the rule and make sure it is ordered first (above other rules).\n\n<Callout type=\"warning\">\n  Also check **Security > Bots** in your Cloudflare dashboard. If \"Bot Fight Mode\" or \"Super Bot Fight Mode\" is enabled, that is very likely what is blocking server-to-server requests from AI tools.\n</Callout>\n\n## Best practices\n\n- **Keep content up-to-date**: Rebuild or redeploy your site after updating documentation to regenerate the index with fresh content.\n- **Use descriptive queries**: More specific queries yield better semantic search results.\n- **Monitor index status**: Use the GET endpoint to check if your index is ready before performing searches.\n- **Optimize content structure**: Well-structured markdown with clear headings improves search relevance.";
+export declare const mcpMdxTemplate = "---\ntitle: \"Model Context Protocol\"\ndescription: \"Connect your Doccupine documentation to AI tools with an MCP server for enhanced AI-powered documentation search.\"\ndate: \"2026-02-19\"\ncategory: \"Configuration\"\ncategoryOrder: 3\norder: 9\n---\n\n# Model Context Protocol\n\nConnect your documentation to AI tools with a hosted MCP server.\n\nDoccupine automatically generates a Model Context Protocol (MCP) server from your documentation, making your content accessible to AI applications like Claude, Cursor, VS Code, and other MCP-compatible tools. Your MCP server exposes semantic search capabilities, allowing AI tools to query your documentation directly and provide accurate, context-aware answers.\n\n<Callout type=\"warning\">\n  The MCP server requires the [AI Assistant](/ai-assistant) to be configured. Make sure you have set up your LLM provider and API keys before using the MCP server.\n</Callout>\n\n## About MCP servers\n\nThe Model Context Protocol (MCP) is an open protocol that creates standardized connections between AI applications and external services, like documentation. Doccupine generates an MCP server from your documentation, preparing your content for the broader AI ecosystem where any MCP client can connect to your documentation.\n\nYour MCP server exposes search and retrieval tools for AI applications to query your documentation. Your users must connect your MCP server to their preferred AI tools to access your documentation.\n\n### How MCP servers work\n\nWhen an AI tool has your documentation MCP server connected, the AI tool can search your documentation directly instead of making a generic web search in response to a user's prompt. Your MCP server provides access to all indexed content from your documentation site.\n\n- The LLM can proactively search your documentation while generating a response, not just when explicitly asked.\n- The LLM determines when to use the search tool based on the context of the conversation and the relevance of your documentation.\n- Each tool call happens during the generation process, so the LLM searches up-to-date information from your documentation to generate its response.\n\n### MCP compared to web search\n\nAI tools can search the web, but MCP provides distinct advantages for documentation.\n\n- **Direct source access**: Web search depends on what search engines have indexed, which may be stale or incomplete. MCP searches your current indexed documentation directly.\n- **Integrated workflow**: MCP allows the AI to search during response generation rather than performing a separate web search.\n- **Semantic search**: MCP uses vector embeddings for semantic similarity search, providing more relevant results than keyword-based web search.\n- **No search noise**: SEO and ranking algorithms influence web search results. MCP goes straight to your documentation content.\n\n## Access your MCP server\n\nDoccupine automatically generates an MCP server for your documentation and hosts it at your documentation URL with the `/api/mcp` path. For example, if your documentation is hosted at `https://example.com`, your MCP server is available at `https://example.com/api/mcp`.\n\nThe MCP server provides both a GET endpoint to discover available tools and a POST endpoint to execute tool calls.\n\n### Authentication\n\nYou can optionally protect your MCP server with an API key by setting the `DOCS_API_KEY` environment variable in your `.env` file:\n\n```bash\nDOCS_API_KEY=your-secret-key\n```\n\nWhen `DOCS_API_KEY` is set, all requests to `/api/mcp` must include an `Authorization` header with a Bearer token:\n\n```text\nAuthorization: Bearer your-secret-key\n```\n\nRequests without a valid token receive a `401 Unauthorized` response. When `DOCS_API_KEY` is not set, the MCP server is publicly accessible with no authentication required.\n\n<Callout type=\"note\">\n  Authentication only applies to the `/api/mcp` endpoint. The `/api/rag` endpoint used by the built-in AI Assistant chat is not affected by this setting.\n</Callout>\n\n### API Endpoints\n\n#### GET /api/mcp\n\nReturns information about available tools and the current index status.\n\n**Response:**\n\n```json\n{\n  \"tools\": [\n    {\n      \"name\": \"search_docs\",\n      \"description\": \"Search through the documentation content using semantic search...\",\n      \"inputSchema\": { ... }\n    },\n    ...\n  ],\n  \"index\": {\n    \"ready\": true,\n    \"chunkCount\": 150\n  }\n}\n```\n\n#### POST /api/mcp\n\nExecutes an MCP tool call.\n\n**Request Body:**\n\n```json\n{\n  \"tool\": \"search_docs\",\n  \"params\": {\n    \"query\": \"how to deploy\",\n    \"limit\": 6\n  }\n}\n```\n\n**Response:**\n\n```json\n{\n  \"content\": [\n    {\n      \"path\": \"app/deployment-and-hosting/page.tsx\",\n      \"uri\": \"docs://deployment-and-hosting\",\n      \"score\": \"0.892\",\n      \"text\": \"Deploy your Doccupine site as a Next.js application...\"\n    },\n    ...\n  ]\n}\n```\n\n## Available Tools\n\nYour MCP server exposes three tools for interacting with your documentation:\n\n### search_docs\n\nSearch through the documentation content using semantic search. Returns relevant chunks of documentation based on the query using vector embeddings and cosine similarity.\n\n**Parameters:**\n\n- `query` (required): The search query to find relevant documentation\n- `limit` (optional): Maximum number of results to return (default: 6)\n\n**Example:**\n\n```json\n{\n  \"tool\": \"search_docs\",\n  \"params\": {\n    \"query\": \"how to configure AI assistant\",\n    \"limit\": 5\n  }\n}\n```\n\n### get_doc\n\nGet the full content of a specific documentation page by its path.\n\n**Parameters:**\n\n- `path` (required): The file path to the documentation page (e.g., `app/getting-started/page.tsx`)\n\n**Example:**\n\n```json\n{\n  \"tool\": \"get_doc\",\n  \"params\": {\n    \"path\": \"app/deployment-and-hosting/page.tsx\"\n  }\n}\n```\n\n### list_docs\n\nList all available documentation pages, optionally filtered by directory.\n\n**Parameters:**\n\n- `directory` (optional): Optional directory to filter results (e.g., `components`)\n\n**Example:**\n\n```json\n{\n  \"tool\": \"list_docs\",\n  \"params\": {\n    \"directory\": \"configuration\"\n  }\n}\n```\n\n## How it works\n\nDoccupine's MCP server uses semantic search powered by vector embeddings to provide accurate, context-aware search results.\n\n### Indexing Process\n\nDoccupine runs this pipeline at build time and ships the resulting vectors with your site, so the server loads a ready-made index instead of re-embedding your docs on every cold start.\n\n1. **Document Discovery**: Doccupine scans your `app/` directory for all `page.tsx`, `page.ts`, `page.jsx`, and `page.js` files.\n2. **Content Extraction**: It extracts content from `const content =` declarations in your page files.\n3. **Chunking**: Large documents are split into chunks of approximately 800 characters with 100 character overlap for better context preservation.\n4. **Embedding Generation**: Each chunk is converted to a vector embedding using your configured LLM provider's embedding model.\n5. **Compaction**: Each vector is truncated to `LLM_EMBEDDING_DIMS` dimensions (default 512) and quantized to int8, keeping the index roughly 20x smaller than raw floats so large doc sets don't stall the chat on cold start.\n6. **Index Building**: The compact embeddings are written to `services/mcp/docs-index.json`, bundled into your serverless functions, and loaded into memory at runtime for fast similarity search.\n\n### Search Process\n\n1. **Query Embedding**: The search query is converted to a vector embedding using the same embedding model.\n2. **Similarity Calculation**: Cosine similarity is calculated between the query embedding and all document chunk embeddings.\n3. **Ranking**: Results are sorted by similarity score (highest first).\n4. **Response**: The top N results (based on the limit parameter) are returned with their paths, URIs, scores, and text content.\n\n<Callout type=\"note\">\n  Embeddings are precomputed at build time and bundled with your site, so the server loads the ready-made index into memory on startup instead of re-embedding your docs on every cold start. It only embeds at runtime when the precomputed index is missing or was built with a different provider or model. When you update your documentation, rebuild or redeploy your site to regenerate the index with fresh content.\n</Callout>\n\n## Use your MCP server\n\nYour users must connect your MCP server to their preferred AI tools.\n\n1. Make your MCP server URL publicly available.\n2. Users copy your MCP server URL and add it to their tools.\n3. Users access your documentation through their tools.\n\nThese are some of the ways you can help your users connect to your MCP server:\n\n<Tabs>\n  <TabContent title=\"Claude\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the steps to connect it to Claude.\n        1. Navigate to the [Connectors](https://claude.ai/settings/connectors) page in the Claude settings.\n        2. Select **Add custom connector**.\n        3. Add your MCP server name and URL.\n        4. Select **Add**.\n        5. When using Claude, select the attachments button (the plus icon).\n        6. Select your MCP server.\n      </Step>\n    </Steps>\n    See the [Model Context Protocol documentation](https://modelcontextprotocol.io/docs/tutorials/use-remote-mcp-server#connecting-to-a-remote-mcp-server) for more details.\n  </TabContent>\n  <TabContent title=\"Claude Code\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the command to connect it to Claude Code.\n        ```bash\n        claude mcp add --transport http <name> <url>\n        ```\n      </Step>\n    </Steps>\n    See the [Claude Code documentation](https://docs.anthropic.com/en/docs/claude-code/mcp#installing-mcp-servers) for more details.\n  </TabContent>\n  <TabContent title=\"Cursor\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the steps to connect it to Cursor.\n        1. Use <kbd>Command</kbd> + <kbd>Shift</kbd> + <kbd>P</kbd> (<kbd>Ctrl</kbd> + <kbd>Shift</kbd> + <kbd>P</kbd> on Windows) to open the command palette.\n        2. Search for \"Open MCP settings\".\n        3. Select **Add custom MCP**. This opens the `mcp.json` file.\n        4. In `mcp.json`, configure your server:\n        ```json\n        {\n          \"mcpServers\": {\n            \"<your-mcp-server-name>\": {\n              \"url\": \"<your-mcp-server-url>\"\n            }\n          }\n        }\n        ```\n      </Step>\n    </Steps>\n    See the [Cursor documentation](https://docs.cursor.com/en/context/mcp#installing-mcp-servers) for more details.\n  </TabContent>\n  <TabContent title=\"VS Code\">\n    <Steps>\n      <Step title=\"Get your MCP server URL.\">\n        Your MCP server URL is available at `https://your-domain.com/api/mcp`.\n      </Step>\n      <Step title=\"Publish your MCP server URL for your users.\">\n        Create a guide for your users that includes your MCP server URL and the steps to connect it to VS Code.\n        1. Create a `.vscode/mcp.json` file.\n        2. In `mcp.json`, configure your server:\n        ```json\n        {\n          \"servers\": {\n            \"<your-mcp-server-name>\": {\n              \"type\": \"http\",\n              \"url\": \"<your-mcp-server-url>\"\n            }\n          }\n        }\n        ```\n      </Step>\n    </Steps>\n    See the [VS Code documentation](https://code.visualstudio.com/docs/copilot/chat/mcp-servers) for more details.\n  </TabContent>\n</Tabs>\n\n## Requirements\n\nTo use the MCP server, you need to have the AI Assistant configured. The MCP server uses the same LLM configuration for generating embeddings.\n\n| Variable       | Required | Description                                                  |\n| -------------- | -------- | ------------------------------------------------------------ |\n| `LLM_PROVIDER` | Yes      | Your LLM provider (`openai`, `anthropic`, or `google`)       |\n| `DOCS_API_KEY` | No       | When set, requires Bearer token authentication on `/api/mcp` |\n\n<Callout type=\"warning\">\n  The MCP server requires an LLM provider to be configured for generating embeddings. Make sure you have set up your AI Assistant with a valid API key before using the MCP server.\n</Callout>\n\nSee the [AI Assistant documentation](/ai-assistant) for configuration details.\n\n## Content indexing\n\nYour MCP server searches content extracted from your page files. The server automatically discovers and indexes all `page.tsx`, `page.ts`, `page.jsx`, and `page.js` files in your `app/` directory.\n\n### Content extraction\n\nThe server extracts content from `const content =` declarations in your page files. Make sure your documentation pages export a `content` constant with your markdown or MDX content.\n\n**Example:**\n\n```tsx\nexport const content = `\n# Getting Started\n\nWelcome to the documentation...\n`;\n```\n\n### Excluded directories\n\nThe following directories are automatically excluded from indexing:\n\n- `node_modules`\n- `.next`\n- `.git`\n- `api`\n\n## Troubleshooting\n\n### Index not building\n\nIf the index is not building, check:\n\n- Your LLM provider is configured correctly in your `.env` file\n- You have a valid API key set\n- Your documentation pages export a `content` constant\n\n### No search results\n\nIf searches return no results:\n\n- Verify that your documentation pages are in the `app/` directory\n- Check that your pages export a `content` constant\n- Ensure the index has been built (check the `index.ready` status via GET `/api/mcp`)\n\n### Slow search performance\n\nSearches use the precomputed in-memory index and are fast. The first search is only slow when the app has to embed your docs at runtime, which happens when the precomputed index is missing or was built with a different provider or model. If performance is consistently slow:\n\n- Confirm the build ran the doc-index precompute step (the `build` script runs it before `next build`) with a valid API key\n- Check your embedding API response times\n- Consider reducing the number of documentation pages\n- Verify your server has sufficient memory\n\n### Cloudflare blocking MCP requests\n\nIf you use Cloudflare as a proxy in front of your documentation site, Cloudflare's bot protection may block server-to-server requests from AI tools like Claude.ai. This can cause MCP connections to fail silently or return errors.\n\nThere are two ways to resolve this:\n\n**Option 1: Disable the Cloudflare proxy (simplest)**\n\nIn your Cloudflare DNS settings, click the orange cloud icon next to your domain record to switch it to \"DNS only\" (grey cloud). This disables Cloudflare's proxy and bot protection for your domain, allowing MCP requests to reach your server directly.\n\n**Option 2: Add a Cloudflare WAF exception (keeps your custom domain proxied)**\n\nIn Cloudflare dashboard:\n\n1. Go to **Security > WAF**.\n2. Click **Create rule**.\n3. Set it up as:\n   - **Rule name**: Allow MCP API\n   - **Field**: URI Path\n   - **Operator**: starts with\n   - **Value**: `/api/mcp`\n   - **Action**: Skip -- then check all remaining custom rules, Rate limiting rules, and Bot Fight Mode / Super Bot Fight Mode.\n4. Deploy the rule and make sure it is ordered first (above other rules).\n\n<Callout type=\"warning\">\n  Also check **Security > Bots** in your Cloudflare dashboard. If \"Bot Fight Mode\" or \"Super Bot Fight Mode\" is enabled, that is very likely what is blocking server-to-server requests from AI tools.\n</Callout>\n\n## Best practices\n\n- **Keep content up-to-date**: Rebuild or redeploy your site after updating documentation to regenerate the index with fresh content.\n- **Use descriptive queries**: More specific queries yield better semantic search results.\n- **Monitor index status**: Use the GET endpoint to check if your index is ready before performing searches.\n- **Optimize content structure**: Well-structured markdown with clear headings improves search relevance.";

package/dist/templates/mdx/model-context-protocol.mdx.js

@@ -198,7 +198,8 @@ Doccupine runs this pipeline at build time and ships the resulting vectors with
 2. **Content Extraction**: It extracts content from \`const content =\` declarations in your page files.
 3. **Chunking**: Large documents are split into chunks of approximately 800 characters with 100 character overlap for better context preservation.
 4. **Embedding Generation**: Each chunk is converted to a vector embedding using your configured LLM provider's embedding model.
-5. **Index Building**: The embeddings are written to \`services/mcp/docs-index.json\`, bundled into your serverless functions, and loaded into memory at runtime for fast similarity search.
+5. **Compaction**: Each vector is truncated to \`LLM_EMBEDDING_DIMS\` dimensions (default 512) and quantized to int8, keeping the index roughly 20x smaller than raw floats so large doc sets don't stall the chat on cold start.
+6. **Index Building**: The compact embeddings are written to \`services/mcp/docs-index.json\`, bundled into your serverless functions, and loaded into memory at runtime for fast similarity search.
 
 ### Search Process
 

package/dist/templates/scripts/build-docs-index.d.ts

@@ -1 +1 @@
-export declare const buildDocsIndexScriptTemplate = "/**\n * Precompute document embeddings at build time.\n *\n * Runs before `next build` (wired into the \"build\" script in package.json).\n * Embeds every docs chunk once and writes services/mcp/docs-index.json, so the\n * running app loads vectors instead of re-embedding the whole doc set on every\n * serverless cold start (the main cause of slow first chats / proxy timeouts).\n *\n * Fails soft: without an API key, or on any embedding error, it leaves the\n * existing (possibly empty) index in place and exits 0 so the build proceeds.\n * The app then falls back to embedding on demand at runtime.\n */\nimport path from \"node:path\";\nimport { writeFileSync } from \"node:fs\";\n// @next/env is CommonJS; use a default import so the named export resolves\n// under tsx's ESM loader (a named import is not statically detected).\nimport nextEnv from \"@next/env\";\nimport { getAllDocsChunks } from \"../services/mcp/tools\";\nimport { getLLMConfig, isLLMAvailable } from \"../services/llm/config\";\nimport { createEmbeddings } from \"../services/llm/factory\";\n\n// This runs as a standalone process before `next build`, so it must load\n// .env / .env.local / .env.production itself (the same way Next does). Real\n// environment variables still take precedence over .env files.\nnextEnv.loadEnvConfig(process.cwd());\n\nconst OUTPUT = path.join(process.cwd(), \"services\", \"mcp\", \"docs-index.json\");\nconst BATCH_SIZE = 10;\n\nasync function main() {\n  if (!isLLMAvailable()) {\n    console.warn(\n      \"[doccupine] No LLM API key set - skipping embedding precompute. \" +\n        \"The chat will embed docs on demand at runtime.\",\n    );\n    return;\n  }\n\n  const chunks = await getAllDocsChunks();\n  if (chunks.length === 0) {\n    console.warn(\"[doccupine] No docs found to embed - skipping precompute.\");\n    return;\n  }\n\n  const config = getLLMConfig();\n  const embeddings = createEmbeddings(config);\n\n  // Embed in small batches to stay within provider token limits.\n  const texts = chunks.map((c) => c.text);\n  const vectors: number[][] = [];\n  for (let i = 0; i < texts.length; i += BATCH_SIZE) {\n    const batch = await embeddings.embedDocuments(\n      texts.slice(i, i + BATCH_SIZE),\n    );\n    vectors.push(...batch);\n  }\n\n  const data = {\n    provider: config.provider,\n    embeddingModel: config.embeddingModel,\n    chunks: chunks.map((c, i) => ({ ...c, embedding: vectors[i] })),\n  };\n\n  writeFileSync(OUTPUT, JSON.stringify(data));\n  console.log(\n    `[doccupine] Precomputed ${data.chunks.length} doc embeddings -> ${OUTPUT}`,\n  );\n}\n\nmain()\n  .then(() => process.exit(0))\n  .catch((error) => {\n    // Never fail the build on an embedding error - fall back to runtime embedding.\n    console.warn(\n      \"[doccupine] Embedding precompute failed; continuing build. \" +\n        \"The chat will embed docs at runtime.\",\n      error instanceof Error ? error.message : error,\n    );\n    process.exit(0);\n  });\n";
+export declare const buildDocsIndexScriptTemplate = "/**\n * Precompute document embeddings at build time.\n *\n * Runs before `next build` (wired into the \"build\" script in package.json).\n * Embeds every docs chunk once and writes services/mcp/docs-index.json, so the\n * running app loads vectors instead of re-embedding the whole doc set on every\n * serverless cold start (the main cause of slow first chats / proxy timeouts).\n *\n * Fails soft: without an API key, or on any embedding error, it leaves the\n * existing (possibly empty) index in place and exits 0 so the build proceeds.\n * The app then falls back to embedding on demand at runtime.\n */\nimport path from \"node:path\";\nimport { writeFileSync } from \"node:fs\";\n// @next/env is CommonJS; use a default import so the named export resolves\n// under tsx's ESM loader (a named import is not statically detected).\nimport nextEnv from \"@next/env\";\nimport { getAllDocsChunks } from \"../services/mcp/tools\";\nimport { getLLMConfig, isLLMAvailable } from \"../services/llm/config\";\nimport { createEmbeddings } from \"../services/llm/factory\";\nimport { reduceDims, quantizeInt8, encodeInt8 } from \"../services/mcp/vector\";\n\n// This runs as a standalone process before `next build`, so it must load\n// .env / .env.local / .env.production itself (the same way Next does). Real\n// environment variables still take precedence over .env files.\nnextEnv.loadEnvConfig(process.cwd());\n\nconst OUTPUT = path.join(process.cwd(), \"services\", \"mcp\", \"docs-index.json\");\nconst BATCH_SIZE = 10;\n\nasync function main() {\n  if (!isLLMAvailable()) {\n    console.warn(\n      \"[doccupine] No LLM API key set - skipping embedding precompute. \" +\n        \"The chat will embed docs on demand at runtime.\",\n    );\n    return;\n  }\n\n  const chunks = await getAllDocsChunks();\n  if (chunks.length === 0) {\n    console.warn(\"[doccupine] No docs found to embed - skipping precompute.\");\n    return;\n  }\n\n  const config = getLLMConfig();\n  const embeddings = createEmbeddings(config);\n\n  // Embed in small batches to stay within provider token limits, then reduce\n  // each vector to config.embeddingDims and quantize to int8. Storing raw float\n  // arrays as JSON balloons to 100MB+ on large doc sets, which OOMs / stalls the\n  // serverless chat on cold start; int8 base64 keeps the index ~20x smaller.\n  const texts = chunks.map((c) => c.text);\n  const encoded: string[] = [];\n  for (let i = 0; i < texts.length; i += BATCH_SIZE) {\n    const batch = await embeddings.embedDocuments(\n      texts.slice(i, i + BATCH_SIZE),\n    );\n    for (const vector of batch) {\n      encoded.push(\n        encodeInt8(quantizeInt8(reduceDims(vector, config.embeddingDims))),\n      );\n    }\n  }\n\n  const data = {\n    provider: config.provider,\n    embeddingModel: config.embeddingModel,\n    dims: config.embeddingDims,\n    quantization: \"int8\",\n    chunks: chunks.map((c, i) => ({ ...c, embedding: encoded[i] })),\n  };\n\n  writeFileSync(OUTPUT, JSON.stringify(data));\n  console.log(\n    `[doccupine] Precomputed ${data.chunks.length} doc embeddings -> ${OUTPUT}`,\n  );\n}\n\nmain()\n  .then(() => process.exit(0))\n  .catch((error) => {\n    // Never fail the build on an embedding error - fall back to runtime embedding.\n    console.warn(\n      \"[doccupine] Embedding precompute failed; continuing build. \" +\n        \"The chat will embed docs at runtime.\",\n      error instanceof Error ? error.message : error,\n    );\n    process.exit(0);\n  });\n";

package/dist/templates/scripts/build-docs-index.js

@@ -18,6 +18,7 @@ import nextEnv from "@next/env";
 import { getAllDocsChunks } from "../services/mcp/tools";
 import { getLLMConfig, isLLMAvailable } from "../services/llm/config";
 import { createEmbeddings } from "../services/llm/factory";
+import { reduceDims, quantizeInt8, encodeInt8 } from "../services/mcp/vector";
 
 // This runs as a standalone process before \`next build\`, so it must load
 // .env / .env.local / .env.production itself (the same way Next does). Real
@@ -45,20 +46,29 @@ async function main() {
   const config = getLLMConfig();
   const embeddings = createEmbeddings(config);
 
-  // Embed in small batches to stay within provider token limits.
+  // Embed in small batches to stay within provider token limits, then reduce
+  // each vector to config.embeddingDims and quantize to int8. Storing raw float
+  // arrays as JSON balloons to 100MB+ on large doc sets, which OOMs / stalls the
+  // serverless chat on cold start; int8 base64 keeps the index ~20x smaller.
   const texts = chunks.map((c) => c.text);
-  const vectors: number[][] = [];
+  const encoded: string[] = [];
   for (let i = 0; i < texts.length; i += BATCH_SIZE) {
     const batch = await embeddings.embedDocuments(
       texts.slice(i, i + BATCH_SIZE),
     );
-    vectors.push(...batch);
+    for (const vector of batch) {
+      encoded.push(
+        encodeInt8(quantizeInt8(reduceDims(vector, config.embeddingDims))),
+      );
+    }
   }
 
   const data = {
     provider: config.provider,
     embeddingModel: config.embeddingModel,
-    chunks: chunks.map((c, i) => ({ ...c, embedding: vectors[i] })),
+    dims: config.embeddingDims,
+    quantization: "int8",
+    chunks: chunks.map((c, i) => ({ ...c, embedding: encoded[i] })),
   };
 
   writeFileSync(OUTPUT, JSON.stringify(data));

package/dist/templates/services/llm/config.d.ts

@@ -1 +1 @@
-export declare const llmConfigTemplate = "import type {\n  LLMConfig,\n  LLMProvider,\n  ProviderDefaults,\n} from \"@/services/llm/types\";\nconst PROVIDER_DEFAULTS: ProviderDefaults = {\n  openai: {\n    chat: \"gpt-4.1-nano\",\n    embedding: \"text-embedding-3-small\",\n  },\n  anthropic: {\n    chat: \"claude-sonnet-4-5-20250929\",\n    embedding: \"text-embedding-3-small\", // Fallback to OpenAI\n  },\n  google: {\n    chat: \"gemini-2.5-flash-lite\",\n    embedding: \"gemini-embedding-001\",\n  },\n};\nfunction validateAPIKeys(provider: LLMProvider): void {\n  const requiredKeys: Record<LLMProvider, string> = {\n    openai: \"OPENAI_API_KEY\",\n    anthropic: \"ANTHROPIC_API_KEY\",\n    google: \"GOOGLE_API_KEY\",\n  };\n  const keyName = requiredKeys[provider];\n  const keyValue = process.env[keyName];\n  if (!keyValue) {\n    throw new Error(\n      `Missing API key for ${provider}. Please set ${keyName} in your environment variables.`,\n    );\n  }\n  if (provider === \"anthropic\" && !process.env.OPENAI_API_KEY) {\n    throw new Error(\n      \"Anthropic provider requires OPENAI_API_KEY for embeddings. Please set OPENAI_API_KEY in your environment variables.\",\n    );\n  }\n}\nexport function isLLMAvailable(): boolean {\n  const provider = (process.env.LLM_PROVIDER || \"openai\") as LLMProvider;\n  const requiredKeys: Record<LLMProvider, string> = {\n    openai: \"OPENAI_API_KEY\",\n    anthropic: \"ANTHROPIC_API_KEY\",\n    google: \"GOOGLE_API_KEY\",\n  };\n  const keyName = requiredKeys[provider];\n  if (!keyName || !process.env[keyName]) return false;\n  if (provider === \"anthropic\" && !process.env.OPENAI_API_KEY) return false;\n  return true;\n}\nexport function getLLMConfig(): LLMConfig {\n  const provider = (process.env.LLM_PROVIDER || \"openai\") as LLMProvider;\n  if (![\"openai\", \"anthropic\", \"google\"].includes(provider)) {\n    throw new Error(\n      `Invalid LLM_PROVIDER: ${provider}. Must be one of: openai, anthropic, google`,\n    );\n  }\n  validateAPIKeys(provider);\n  const defaults = PROVIDER_DEFAULTS[provider];\n  return {\n    provider,\n    chatModel: process.env.LLM_CHAT_MODEL || defaults.chat,\n    embeddingModel: process.env.LLM_EMBEDDING_MODEL || defaults.embedding,\n    temperature: parseFloat(process.env.LLM_TEMPERATURE || \"0\"),\n  };\n}\n";
+export declare const llmConfigTemplate = "import type {\n  LLMConfig,\n  LLMProvider,\n  ProviderDefaults,\n} from \"@/services/llm/types\";\nconst PROVIDER_DEFAULTS: ProviderDefaults = {\n  openai: {\n    chat: \"gpt-4.1-nano\",\n    embedding: \"text-embedding-3-small\",\n  },\n  anthropic: {\n    chat: \"claude-sonnet-4-5-20250929\",\n    embedding: \"text-embedding-3-small\", // Fallback to OpenAI\n  },\n  google: {\n    chat: \"gemini-2.5-flash-lite\",\n    embedding: \"gemini-embedding-001\",\n  },\n};\nfunction validateAPIKeys(provider: LLMProvider): void {\n  const requiredKeys: Record<LLMProvider, string> = {\n    openai: \"OPENAI_API_KEY\",\n    anthropic: \"ANTHROPIC_API_KEY\",\n    google: \"GOOGLE_API_KEY\",\n  };\n  const keyName = requiredKeys[provider];\n  const keyValue = process.env[keyName];\n  if (!keyValue) {\n    throw new Error(\n      `Missing API key for ${provider}. Please set ${keyName} in your environment variables.`,\n    );\n  }\n  if (provider === \"anthropic\" && !process.env.OPENAI_API_KEY) {\n    throw new Error(\n      \"Anthropic provider requires OPENAI_API_KEY for embeddings. Please set OPENAI_API_KEY in your environment variables.\",\n    );\n  }\n}\nexport function isLLMAvailable(): boolean {\n  const provider = (process.env.LLM_PROVIDER || \"openai\") as LLMProvider;\n  const requiredKeys: Record<LLMProvider, string> = {\n    openai: \"OPENAI_API_KEY\",\n    anthropic: \"ANTHROPIC_API_KEY\",\n    google: \"GOOGLE_API_KEY\",\n  };\n  const keyName = requiredKeys[provider];\n  if (!keyName || !process.env[keyName]) return false;\n  if (provider === \"anthropic\" && !process.env.OPENAI_API_KEY) return false;\n  return true;\n}\nexport function getLLMConfig(): LLMConfig {\n  const provider = (process.env.LLM_PROVIDER || \"openai\") as LLMProvider;\n  if (![\"openai\", \"anthropic\", \"google\"].includes(provider)) {\n    throw new Error(\n      `Invalid LLM_PROVIDER: ${provider}. Must be one of: openai, anthropic, google`,\n    );\n  }\n  validateAPIKeys(provider);\n  const defaults = PROVIDER_DEFAULTS[provider];\n  // Vectors are Matryoshka-truncated to this many dimensions before being\n  // stored as int8, keeping the prebuilt search index small. Must match between\n  // build time and runtime; a mismatch forces a re-embed. Values >= the model's\n  // native dimension keep full precision (default: 512).\n  const rawDims = process.env.LLM_EMBEDDING_DIMS;\n  let embeddingDims = 512;\n  if (rawDims !== undefined && rawDims !== \"\") {\n    const parsed = parseInt(rawDims, 10);\n    if (Number.isFinite(parsed) && parsed >= 0) embeddingDims = parsed;\n  }\n  return {\n    provider,\n    chatModel: process.env.LLM_CHAT_MODEL || defaults.chat,\n    embeddingModel: process.env.LLM_EMBEDDING_MODEL || defaults.embedding,\n    embeddingDims,\n    temperature: parseFloat(process.env.LLM_TEMPERATURE || \"0\"),\n  };\n}\n";

package/dist/templates/services/llm/config.js

@@ -57,10 +57,21 @@ export function getLLMConfig(): LLMConfig {
   }
   validateAPIKeys(provider);
   const defaults = PROVIDER_DEFAULTS[provider];
+  // Vectors are Matryoshka-truncated to this many dimensions before being
+  // stored as int8, keeping the prebuilt search index small. Must match between
+  // build time and runtime; a mismatch forces a re-embed. Values >= the model's
+  // native dimension keep full precision (default: 512).
+  const rawDims = process.env.LLM_EMBEDDING_DIMS;
+  let embeddingDims = 512;
+  if (rawDims !== undefined && rawDims !== "") {
+    const parsed = parseInt(rawDims, 10);
+    if (Number.isFinite(parsed) && parsed >= 0) embeddingDims = parsed;
+  }
   return {
     provider,
     chatModel: process.env.LLM_CHAT_MODEL || defaults.chat,
     embeddingModel: process.env.LLM_EMBEDDING_MODEL || defaults.embedding,
+    embeddingDims,
     temperature: parseFloat(process.env.LLM_TEMPERATURE || "0"),
   };
 }

package/dist/templates/services/llm/types.d.ts

@@ -1 +1 @@
-export declare const llmTypesTemplate = "export type LLMProvider = \"openai\" | \"anthropic\" | \"google\";\n\nexport interface LLMConfig {\n  provider: LLMProvider;\n  chatModel: string;\n  embeddingModel: string;\n  temperature: number;\n}\n\ninterface ProviderModels {\n  chat: string;\n  embedding: string;\n}\n\nexport type ProviderDefaults = Record<LLMProvider, ProviderModels>;\n";
+export declare const llmTypesTemplate = "export type LLMProvider = \"openai\" | \"anthropic\" | \"google\";\n\nexport interface LLMConfig {\n  provider: LLMProvider;\n  chatModel: string;\n  embeddingModel: string;\n  embeddingDims: number;\n  temperature: number;\n}\n\ninterface ProviderModels {\n  chat: string;\n  embedding: string;\n}\n\nexport type ProviderDefaults = Record<LLMProvider, ProviderModels>;\n";

package/dist/templates/services/llm/types.js

@@ -4,6 +4,7 @@ export interface LLMConfig {
   provider: LLMProvider;
   chatModel: string;
   embeddingModel: string;
+  embeddingDims: number;
   temperature: number;
 }
 

package/dist/templates/services/mcp/docsIndexStub.d.ts

@@ -1 +1 @@
-export declare const docsIndexStubTemplate = "{\n  \"provider\": null,\n  \"embeddingModel\": null,\n  \"chunks\": []\n}\n";
+export declare const docsIndexStubTemplate = "{\n  \"provider\": null,\n  \"embeddingModel\": null,\n  \"dims\": 0,\n  \"quantization\": \"none\",\n  \"chunks\": []\n}\n";

package/dist/templates/services/mcp/docsIndexStub.js

@@ -4,6 +4,8 @@
 export const docsIndexStubTemplate = `{
   "provider": null,
   "embeddingModel": null,
+  "dims": 0,
+  "quantization": "none",
   "chunks": []
 }
 `;

package/dist/templates/services/mcp/index.d.ts

@@ -1 +1 @@
-export declare const mcpIndexTemplate = "export * from \"@/services/mcp/types\";\nexport * from \"@/services/mcp/tools\";\nexport * from \"@/services/mcp/server\";\n";
+export declare const mcpIndexTemplate = "export * from \"@/services/mcp/types\";\nexport * from \"@/services/mcp/tools\";\nexport * from \"@/services/mcp/vector\";\nexport * from \"@/services/mcp/server\";\n";

package/dist/templates/services/mcp/index.js

@@ -1,4 +1,5 @@
 export const mcpIndexTemplate = `export * from "@/services/mcp/types";
 export * from "@/services/mcp/tools";
+export * from "@/services/mcp/vector";
 export * from "@/services/mcp/server";
 `;

package/dist/templates/services/mcp/server.d.ts

@@ -1 +1 @@
-export declare const mcpServerTemplate = "import path from \"node:path\";\nimport fs from \"node:fs\";\nimport { McpServer } from \"@modelcontextprotocol/sdk/server/mcp.js\";\nimport { z } from \"zod\";\nimport {\n  listDocs,\n  getDoc,\n  getAllDocsChunks,\n  DOCS_TOOLS,\n} from \"@/services/mcp/tools\";\nimport { getLLMConfig, isLLMAvailable, createEmbeddings } from \"@/services/llm\";\nimport type { DocsChunk } from \"@/services/mcp/types\";\n\n/**\n * In-memory cache for document embeddings.\n * Built once at server startup since docs are static.\n */\nlet docsIndex: {\n  ready: boolean;\n  building: boolean;\n  chunks: (DocsChunk & { embedding: number[] })[];\n} = {\n  ready: false,\n  building: false,\n  chunks: [],\n};\n\n/** Resolves when the initial index build completes */\nlet indexReady: Promise<void> | null = null;\n\n/**\n * Cosine similarity between two vectors\n */\nfunction cosineSim(a: number[], b: number[]): number {\n  let dot = 0,\n    na = 0,\n    nb = 0;\n  for (let i = 0; i < a.length; i++) {\n    const x = a[i];\n    const y = b[i];\n    dot += x * y;\n    na += x * x;\n    nb += y * y;\n  }\n  if (na === 0 || nb === 0) return 0;\n  return dot / (Math.sqrt(na) * Math.sqrt(nb));\n}\n\n/**\n * Absolute path to the embeddings index precomputed at build time by\n * scripts/build-docs-index.mts. Bundled into serverless functions via\n * outputFileTracingIncludes in next.config.ts.\n */\nconst INDEX_FILE = path.join(\n  process.cwd(),\n  \"services\",\n  \"mcp\",\n  \"docs-index.json\",\n);\n\n/**\n * Load embeddings precomputed at build time. Returns null when the file is\n * missing/empty or was built with a different provider/model than the current\n * config - query and document vectors must come from the same embedding model.\n */\nfunction loadPrecomputedIndex():\n  (DocsChunk & { embedding: number[] })[] | null {\n  try {\n    const parsed = JSON.parse(fs.readFileSync(INDEX_FILE, \"utf8\")) as {\n      provider?: string;\n      embeddingModel?: string;\n      chunks?: (DocsChunk & { embedding: number[] })[];\n    };\n    if (!parsed.chunks || parsed.chunks.length === 0) return null;\n    const config = getLLMConfig();\n    if (\n      parsed.provider !== config.provider ||\n      parsed.embeddingModel !== config.embeddingModel\n    ) {\n      return null;\n    }\n    return parsed.chunks;\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Build or rebuild the documentation index\n */\nasync function buildDocsIndex(force = false): Promise<void> {\n  if (docsIndex.building) return;\n  if (docsIndex.ready && !force) return;\n\n  docsIndex.building = true;\n  try {\n    // Prefer embeddings precomputed at build time - avoids re-embedding the\n    // entire doc set on every cold start (the main cause of slow first chats).\n    if (!force) {\n      const precomputed = loadPrecomputedIndex();\n      if (precomputed) {\n        docsIndex.chunks = precomputed;\n        docsIndex.ready = true;\n        return;\n      }\n    }\n\n    const chunks = await getAllDocsChunks();\n\n    if (chunks.length === 0) {\n      docsIndex.chunks = [];\n      docsIndex.ready = true;\n      return;\n    }\n\n    const config = getLLMConfig();\n    const embeddings = createEmbeddings(config);\n\n    // Process embeddings in small batches to avoid exceeding token limits\n    const BATCH_SIZE = 10;\n    const texts = chunks.map((c) => c.text);\n    const vectors: number[][] = [];\n\n    for (let i = 0; i < texts.length; i += BATCH_SIZE) {\n      const batch = texts.slice(i, i + BATCH_SIZE);\n      const batchVectors = await embeddings.embedDocuments(batch);\n      vectors.push(...batchVectors);\n    }\n\n    docsIndex.chunks = chunks.map((c, i) => ({\n      ...c,\n      embedding: vectors[i],\n    }));\n    docsIndex.ready = true;\n  } catch (error) {\n    // Reset so the next call to ensureDocsIndex retries\n    indexReady = null;\n    throw error;\n  } finally {\n    docsIndex.building = false;\n  }\n}\n\n/**\n * Ensure the docs index is ready.\n * On first call, triggers the build; subsequent calls wait for the same promise.\n */\nexport async function ensureDocsIndex(force = false): Promise<void> {\n  if (force) {\n    // Wait for any in-flight build before starting a forced rebuild\n    if (docsIndex.building && indexReady) {\n      await indexReady.catch(() => {});\n    }\n    docsIndex.ready = false;\n    docsIndex.chunks = [];\n    indexReady = buildDocsIndex(true);\n    return indexReady;\n  }\n  if (!indexReady) {\n    indexReady = buildDocsIndex();\n  }\n  return indexReady;\n}\n\n// Eagerly start building the index on server startup if LLM is configured\nif (isLLMAvailable()) {\n  indexReady = buildDocsIndex();\n}\n\n/** Cached embeddings instance for search queries */\nlet cachedEmbeddings: ReturnType<typeof createEmbeddings> | null = null;\n\nfunction getEmbeddings() {\n  if (!cachedEmbeddings) {\n    cachedEmbeddings = createEmbeddings(getLLMConfig());\n  }\n  return cachedEmbeddings;\n}\n\n/**\n * Search documents using semantic similarity\n */\nexport async function searchDocs(\n  query: string,\n  limit = 6,\n): Promise<{ chunk: DocsChunk; score: number }[]> {\n  await ensureDocsIndex();\n\n  const queryVector = await getEmbeddings().embedQuery(query);\n\n  const scored = docsIndex.chunks\n    .map((c) => ({\n      chunk: { id: c.id, text: c.text, path: c.path, uri: c.uri },\n      score: cosineSim(queryVector, c.embedding),\n    }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, limit);\n\n  return scored;\n}\n\n/**\n * Get the current index status\n */\nexport function getIndexStatus(): { ready: boolean; chunkCount: number } {\n  return {\n    ready: docsIndex.ready,\n    chunkCount: docsIndex.chunks.length,\n  };\n}\n\n/**\n * Create and configure the MCP server with documentation tools\n */\nexport function createMCPServer(): McpServer {\n  const server = new McpServer({\n    name: \"docs-server\",\n    version: \"1.0.0\",\n  });\n\n  // Register the search_docs tool\n  server.tool(\n    \"search_docs\",\n    DOCS_TOOLS[0].description,\n    {\n      query: z\n        .string()\n        .describe(\"The search query to find relevant documentation\"),\n      limit: z\n        .number()\n        .optional()\n        .describe(\"Maximum number of results to return (default: 6)\"),\n    },\n    async ({ query, limit }) => {\n      const results = await searchDocs(query, limit ?? 6);\n      return {\n        content: [\n          {\n            type: \"text\" as const,\n            text: JSON.stringify(\n              results.map(({ chunk, score }) => ({\n                path: chunk.path,\n                uri: chunk.uri,\n                score: score.toFixed(3),\n                text: chunk.text,\n              })),\n              null,\n              2,\n            ),\n          },\n        ],\n      };\n    },\n  );\n\n  // Register the get_doc tool\n  server.tool(\n    \"get_doc\",\n    DOCS_TOOLS[1].description,\n    {\n      path: z.string().describe(\"The file path to the documentation page\"),\n    },\n    async ({ path }) => {\n      const doc = await getDoc({ path });\n      if (!doc) {\n        return {\n          content: [\n            {\n              type: \"text\" as const,\n              text: JSON.stringify({ error: \"Document not found\" }),\n            },\n          ],\n          isError: true,\n        };\n      }\n      return {\n        content: [\n          {\n            type: \"text\" as const,\n            text: JSON.stringify(doc, null, 2),\n          },\n        ],\n      };\n    },\n  );\n\n  // Register the list_docs tool\n  server.tool(\n    \"list_docs\",\n    DOCS_TOOLS[2].description,\n    {\n      directory: z\n        .string()\n        .optional()\n        .describe(\"Optional directory to filter results\"),\n    },\n    async ({ directory }) => {\n      const docs = await listDocs({ directory });\n      return {\n        content: [\n          {\n            type: \"text\" as const,\n            text: JSON.stringify(\n              docs.map((d) => ({\n                name: d.name,\n                path: d.path,\n                uri: d.uri,\n              })),\n              null,\n              2,\n            ),\n          },\n        ],\n      };\n    },\n  );\n\n  // Register documentation as resources\n  server.resource(\"docs://list\", \"docs://list\", async () => {\n    const docs = await listDocs();\n    return {\n      contents: [\n        {\n          uri: \"docs://list\",\n          text: JSON.stringify(\n            docs.map((d) => ({ name: d.name, path: d.path, uri: d.uri })),\n            null,\n            2,\n          ),\n        },\n      ],\n    };\n  });\n\n  return server;\n}\n";
+export declare const mcpServerTemplate = "import path from \"node:path\";\nimport fs from \"node:fs\";\nimport { McpServer } from \"@modelcontextprotocol/sdk/server/mcp.js\";\nimport { z } from \"zod\";\nimport {\n  listDocs,\n  getDoc,\n  getAllDocsChunks,\n  DOCS_TOOLS,\n} from \"@/services/mcp/tools\";\nimport { getLLMConfig, isLLMAvailable, createEmbeddings } from \"@/services/llm\";\nimport {\n  reduceDims,\n  quantizeInt8,\n  decodeInt8,\n  cosineFloatInt8,\n} from \"@/services/mcp/vector\";\nimport type { DocsChunk } from \"@/services/mcp/types\";\n\n/** A doc chunk with its stored int8-quantized embedding. */\ntype IndexedChunk = DocsChunk & { embedding: Int8Array };\n\n/**\n * Thrown when a query arrives but no usable prebuilt embeddings index exists AND\n * the doc set is too large to embed within a single serverless request. Surfaced\n * to the client as a clear \"temporarily unavailable\" message instead of letting\n * the request hang until the platform's function timeout (the old failure mode:\n * the chat connected, streamed heartbeats, and never answered on large docs).\n */\nexport class IndexNotBuiltError extends Error {\n  constructor(message: string) {\n    super(message);\n    this.name = \"IndexNotBuiltError\";\n  }\n}\n\n/**\n * Max chunks we will embed on demand inside one request when the prebuilt index\n * is missing. Above this, embedding the whole set (one round trip per BATCH_SIZE\n * chunks) cannot finish within a serverless function's time limit, so the chat\n * would hang forever. Only enforced in production - `next dev` has no time cap,\n * so it keeps embedding on demand and works without running the build. Override\n * per deployment with RAG_RUNTIME_EMBED_MAX_CHUNKS (0 = always require a prebuilt\n * index). Parsed defensively so a bad value falls back to the default rather\n * than silently disabling the guard; note `Number(x) || 400` would drop a valid 0.\n */\nfunction resolveRuntimeEmbedMax(): number {\n  const raw = process.env.RAG_RUNTIME_EMBED_MAX_CHUNKS;\n  if (raw !== undefined && raw !== \"\") {\n    const parsed = Number(raw);\n    if (Number.isFinite(parsed) && parsed >= 0) return Math.floor(parsed);\n  }\n  return 400;\n}\nconst RUNTIME_EMBED_MAX_CHUNKS = resolveRuntimeEmbedMax();\n\n/**\n * In-memory cache for document embeddings.\n * Built once at server startup since docs are static.\n */\nlet docsIndex: {\n  ready: boolean;\n  building: boolean;\n  chunks: IndexedChunk[];\n} = {\n  ready: false,\n  building: false,\n  chunks: [],\n};\n\n/** Resolves when the initial index build completes */\nlet indexReady: Promise<void> | null = null;\n\n/**\n * Human-readable reason the index is unavailable, or null when healthy. Surfaced\n * by getIndexStatus (and GET /api/rag) so a missing/broken index is diagnosable\n * from outside without reading server logs.\n */\nlet indexUnavailableReason: string | null = null;\n\n/**\n * Absolute path to the embeddings index precomputed at build time by\n * scripts/build-docs-index.mts. Bundled into serverless functions via\n * outputFileTracingIncludes in next.config.ts.\n */\nconst INDEX_FILE = path.join(\n  process.cwd(),\n  \"services\",\n  \"mcp\",\n  \"docs-index.json\",\n);\n\n/**\n * Load embeddings precomputed at build time. Returns null when the file is\n * missing/empty, was built with a different provider/model/dimension than the\n * current config, or is not int8-quantized - query and document vectors must\n * come from the same embedding model and the same transform (dims + int8).\n * A null return makes the caller fall back to embedding on demand at runtime.\n */\nfunction loadPrecomputedIndex(): IndexedChunk[] | null {\n  try {\n    const parsed = JSON.parse(fs.readFileSync(INDEX_FILE, \"utf8\")) as {\n      provider?: string;\n      embeddingModel?: string;\n      dims?: number;\n      quantization?: string;\n      chunks?: (DocsChunk & { embedding: string })[];\n    };\n    if (!parsed.chunks || parsed.chunks.length === 0) return null;\n    const config = getLLMConfig();\n    // Guard on the exact transform: the same dims value at build and query time\n    // guarantees reduceDims produces matching-length vectors on both sides.\n    if (\n      parsed.provider !== config.provider ||\n      parsed.embeddingModel !== config.embeddingModel ||\n      parsed.dims !== config.embeddingDims ||\n      parsed.quantization !== \"int8\"\n    ) {\n      return null;\n    }\n    const decoded: IndexedChunk[] = [];\n    let expectedLen = -1;\n    for (const c of parsed.chunks) {\n      const embedding = decodeInt8(c.embedding);\n      // Reject a corrupt index rather than scoring against ragged vectors.\n      if (expectedLen === -1) expectedLen = embedding.length;\n      else if (embedding.length !== expectedLen) return null;\n      decoded.push({\n        id: c.id,\n        text: c.text,\n        path: c.path,\n        uri: c.uri,\n        embedding,\n      });\n    }\n    return decoded;\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Build or rebuild the documentation index\n */\nasync function buildDocsIndex(force = false): Promise<void> {\n  if (docsIndex.building) return;\n  if (docsIndex.ready && !force) return;\n\n  docsIndex.building = true;\n  indexUnavailableReason = null;\n  try {\n    // Prefer embeddings precomputed at build time - avoids re-embedding the\n    // entire doc set on every cold start (the main cause of slow first chats).\n    if (!force) {\n      const precomputed = loadPrecomputedIndex();\n      if (precomputed) {\n        docsIndex.chunks = precomputed;\n        docsIndex.ready = true;\n        return;\n      }\n    }\n\n    const chunks = await getAllDocsChunks();\n\n    if (chunks.length === 0) {\n      docsIndex.chunks = [];\n      docsIndex.ready = true;\n      return;\n    }\n\n    // Reaching here means no usable prebuilt index was loaded (missing, empty,\n    // or built with a different provider/model/dims). Embedding the whole set on\n    // demand only completes in time for small doc sets; for large ones it runs\n    // one round trip per BATCH_SIZE chunks and blows past the serverless function\n    // limit, so the chat connects and then hangs forever with no answer. In\n    // production, fail fast with an actionable error instead of hanging (this\n    // also stops a client-forced `refresh` from burning embedding quota).\n    if (\n      process.env.NODE_ENV === \"production\" &&\n      chunks.length > RUNTIME_EMBED_MAX_CHUNKS\n    ) {\n      indexUnavailableReason =\n        `Prebuilt embeddings index missing; refusing to embed ${chunks.length} ` +\n        `chunks at request time (limit ${RUNTIME_EMBED_MAX_CHUNKS}). Run the build ` +\n        `with an embedding API key set so scripts/build-docs-index.mts writes ` +\n        `services/mcp/docs-index.json, then redeploy.`;\n      console.error(`[doccupine] ${indexUnavailableReason}`);\n      throw new IndexNotBuiltError(\n        \"The AI assistant is temporarily unavailable: its documentation search \" +\n          \"index has not been built for this deployment. If you are the site \" +\n          \"owner, redeploy with an embedding API key available at build time.\",\n      );\n    }\n\n    // Small enough (or local dev): embed on demand. This still re-embeds on every\n    // cold start, so a prebuilt index is strongly preferred in production.\n    console.warn(\n      `[doccupine] No prebuilt embeddings index found; embedding ${chunks.length} ` +\n        `chunks on demand. Precompute at build time to avoid this on each cold start.`,\n    );\n\n    const config = getLLMConfig();\n    const embeddings = createEmbeddings(config);\n\n    // Process embeddings in small batches to avoid exceeding token limits.\n    // Reduce + quantize to the same int8 representation as the precomputed\n    // index so searchDocs scores identically on either path.\n    const BATCH_SIZE = 10;\n    const texts = chunks.map((c) => c.text);\n    const built: IndexedChunk[] = [];\n\n    for (let i = 0; i < texts.length; i += BATCH_SIZE) {\n      const batch = texts.slice(i, i + BATCH_SIZE);\n      const batchVectors = await embeddings.embedDocuments(batch);\n      for (let j = 0; j < batchVectors.length; j++) {\n        built.push({\n          ...chunks[i + j],\n          embedding: quantizeInt8(\n            reduceDims(batchVectors[j], config.embeddingDims),\n          ),\n        });\n      }\n    }\n\n    docsIndex.chunks = built;\n    docsIndex.ready = true;\n  } catch (error) {\n    // Reset so the next call to ensureDocsIndex retries\n    indexReady = null;\n    throw error;\n  } finally {\n    docsIndex.building = false;\n  }\n}\n\n/**\n * Ensure the docs index is ready.\n * On first call, triggers the build; subsequent calls wait for the same promise.\n */\nexport async function ensureDocsIndex(force = false): Promise<void> {\n  if (force) {\n    // Wait for any in-flight build before starting a forced rebuild\n    if (docsIndex.building && indexReady) {\n      await indexReady.catch(() => {});\n    }\n    docsIndex.ready = false;\n    docsIndex.chunks = [];\n    indexReady = buildDocsIndex(true);\n    return indexReady;\n  }\n  if (!indexReady) {\n    indexReady = buildDocsIndex();\n  }\n  return indexReady;\n}\n\n// Eagerly start building the index on server startup if LLM is configured\nif (isLLMAvailable()) {\n  const initialBuild = buildDocsIndex();\n  indexReady = initialBuild;\n  // A failed eager build (e.g. a missing prebuilt index on a large doc set) must\n  // not surface as an unhandledRejection at startup; the first request retries\n  // via ensureDocsIndex and returns a real error to the client.\n  void initialBuild.catch(() => {});\n}\n\n/** Cached embeddings instance for search queries */\nlet cachedEmbeddings: ReturnType<typeof createEmbeddings> | null = null;\n\nfunction getEmbeddings() {\n  if (!cachedEmbeddings) {\n    cachedEmbeddings = createEmbeddings(getLLMConfig());\n  }\n  return cachedEmbeddings;\n}\n\n/**\n * Search documents using semantic similarity\n */\nexport async function searchDocs(\n  query: string,\n  limit = 6,\n): Promise<{ chunk: DocsChunk; score: number }[]> {\n  await ensureDocsIndex();\n\n  // Reduce the query vector with the exact same transform used at index time so\n  // dimensions line up. Scoring runs directly against the int8 vectors - cosine\n  // is scale-invariant, so no dequantization is needed.\n  const rawQueryVector = await getEmbeddings().embedQuery(query);\n  const queryVector = reduceDims(rawQueryVector, getLLMConfig().embeddingDims);\n\n  const scored = docsIndex.chunks\n    .map((c) => ({\n      chunk: { id: c.id, text: c.text, path: c.path, uri: c.uri },\n      score: cosineFloatInt8(queryVector, c.embedding),\n    }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, limit);\n\n  return scored;\n}\n\n/**\n * Get the current index status\n */\nexport function getIndexStatus(): {\n  ready: boolean;\n  chunkCount: number;\n  reason: string | null;\n} {\n  return {\n    ready: docsIndex.ready,\n    chunkCount: docsIndex.chunks.length,\n    reason: indexUnavailableReason,\n  };\n}\n\n/**\n * Create and configure the MCP server with documentation tools\n */\nexport function createMCPServer(): McpServer {\n  const server = new McpServer({\n    name: \"docs-server\",\n    version: \"1.0.0\",\n  });\n\n  // Register the search_docs tool\n  server.tool(\n    \"search_docs\",\n    DOCS_TOOLS[0].description,\n    {\n      query: z\n        .string()\n        .describe(\"The search query to find relevant documentation\"),\n      limit: z\n        .number()\n        .optional()\n        .describe(\"Maximum number of results to return (default: 6)\"),\n    },\n    async ({ query, limit }) => {\n      const results = await searchDocs(query, limit ?? 6);\n      return {\n        content: [\n          {\n            type: \"text\" as const,\n            text: JSON.stringify(\n              results.map(({ chunk, score }) => ({\n                path: chunk.path,\n                uri: chunk.uri,\n                score: score.toFixed(3),\n                text: chunk.text,\n              })),\n              null,\n              2,\n            ),\n          },\n        ],\n      };\n    },\n  );\n\n  // Register the get_doc tool\n  server.tool(\n    \"get_doc\",\n    DOCS_TOOLS[1].description,\n    {\n      path: z.string().describe(\"The file path to the documentation page\"),\n    },\n    async ({ path }) => {\n      const doc = await getDoc({ path });\n      if (!doc) {\n        return {\n          content: [\n            {\n              type: \"text\" as const,\n              text: JSON.stringify({ error: \"Document not found\" }),\n            },\n          ],\n          isError: true,\n        };\n      }\n      return {\n        content: [\n          {\n            type: \"text\" as const,\n            text: JSON.stringify(doc, null, 2),\n          },\n        ],\n      };\n    },\n  );\n\n  // Register the list_docs tool\n  server.tool(\n    \"list_docs\",\n    DOCS_TOOLS[2].description,\n    {\n      directory: z\n        .string()\n        .optional()\n        .describe(\"Optional directory to filter results\"),\n    },\n    async ({ directory }) => {\n      const docs = await listDocs({ directory });\n      return {\n        content: [\n          {\n            type: \"text\" as const,\n            text: JSON.stringify(\n              docs.map((d) => ({\n                name: d.name,\n                path: d.path,\n                uri: d.uri,\n              })),\n              null,\n              2,\n            ),\n          },\n        ],\n      };\n    },\n  );\n\n  // Register documentation as resources\n  server.resource(\"docs://list\", \"docs://list\", async () => {\n    const docs = await listDocs();\n    return {\n      contents: [\n        {\n          uri: \"docs://list\",\n          text: JSON.stringify(\n            docs.map((d) => ({ name: d.name, path: d.path, uri: d.uri })),\n            null,\n            2,\n          ),\n        },\n      ],\n    };\n  });\n\n  return server;\n}\n";

package/dist/templates/services/mcp/server.js

@@ -9,8 +9,51 @@ import {
   DOCS_TOOLS,
 } from "@/services/mcp/tools";
 import { getLLMConfig, isLLMAvailable, createEmbeddings } from "@/services/llm";
+import {
+  reduceDims,
+  quantizeInt8,
+  decodeInt8,
+  cosineFloatInt8,
+} from "@/services/mcp/vector";
 import type { DocsChunk } from "@/services/mcp/types";
 
+/** A doc chunk with its stored int8-quantized embedding. */
+type IndexedChunk = DocsChunk & { embedding: Int8Array };
+
+/**
+ * Thrown when a query arrives but no usable prebuilt embeddings index exists AND
+ * the doc set is too large to embed within a single serverless request. Surfaced
+ * to the client as a clear "temporarily unavailable" message instead of letting
+ * the request hang until the platform's function timeout (the old failure mode:
+ * the chat connected, streamed heartbeats, and never answered on large docs).
+ */
+export class IndexNotBuiltError extends Error {
+  constructor(message: string) {
+    super(message);
+    this.name = "IndexNotBuiltError";
+  }
+}
+
+/**
+ * Max chunks we will embed on demand inside one request when the prebuilt index
+ * is missing. Above this, embedding the whole set (one round trip per BATCH_SIZE
+ * chunks) cannot finish within a serverless function's time limit, so the chat
+ * would hang forever. Only enforced in production - \`next dev\` has no time cap,
+ * so it keeps embedding on demand and works without running the build. Override
+ * per deployment with RAG_RUNTIME_EMBED_MAX_CHUNKS (0 = always require a prebuilt
+ * index). Parsed defensively so a bad value falls back to the default rather
+ * than silently disabling the guard; note \`Number(x) || 400\` would drop a valid 0.
+ */
+function resolveRuntimeEmbedMax(): number {
+  const raw = process.env.RAG_RUNTIME_EMBED_MAX_CHUNKS;
+  if (raw !== undefined && raw !== "") {
+    const parsed = Number(raw);
+    if (Number.isFinite(parsed) && parsed >= 0) return Math.floor(parsed);
+  }
+  return 400;
+}
+const RUNTIME_EMBED_MAX_CHUNKS = resolveRuntimeEmbedMax();
+
 /**
  * In-memory cache for document embeddings.
  * Built once at server startup since docs are static.
@@ -18,7 +61,7 @@ import type { DocsChunk } from "@/services/mcp/types";
 let docsIndex: {
   ready: boolean;
   building: boolean;
-  chunks: (DocsChunk & { embedding: number[] })[];
+  chunks: IndexedChunk[];
 } = {
   ready: false,
   building: false,
@@ -29,22 +72,11 @@ let docsIndex: {
 let indexReady: Promise<void> | null = null;
 
 /**
- * Cosine similarity between two vectors
+ * Human-readable reason the index is unavailable, or null when healthy. Surfaced
+ * by getIndexStatus (and GET /api/rag) so a missing/broken index is diagnosable
+ * from outside without reading server logs.
  */
-function cosineSim(a: number[], b: number[]): number {
-  let dot = 0,
-    na = 0,
-    nb = 0;
-  for (let i = 0; i < a.length; i++) {
-    const x = a[i];
-    const y = b[i];
-    dot += x * y;
-    na += x * x;
-    nb += y * y;
-  }
-  if (na === 0 || nb === 0) return 0;
-  return dot / (Math.sqrt(na) * Math.sqrt(nb));
-}
+let indexUnavailableReason: string | null = null;
 
 /**
  * Absolute path to the embeddings index precomputed at build time by
@@ -60,26 +92,48 @@ const INDEX_FILE = path.join(
 
 /**
  * Load embeddings precomputed at build time. Returns null when the file is
- * missing/empty or was built with a different provider/model than the current
- * config - query and document vectors must come from the same embedding model.
+ * missing/empty, was built with a different provider/model/dimension than the
+ * current config, or is not int8-quantized - query and document vectors must
+ * come from the same embedding model and the same transform (dims + int8).
+ * A null return makes the caller fall back to embedding on demand at runtime.
  */
-function loadPrecomputedIndex():
-  (DocsChunk & { embedding: number[] })[] | null {
+function loadPrecomputedIndex(): IndexedChunk[] | null {
   try {
     const parsed = JSON.parse(fs.readFileSync(INDEX_FILE, "utf8")) as {
       provider?: string;
       embeddingModel?: string;
-      chunks?: (DocsChunk & { embedding: number[] })[];
+      dims?: number;
+      quantization?: string;
+      chunks?: (DocsChunk & { embedding: string })[];
     };
     if (!parsed.chunks || parsed.chunks.length === 0) return null;
     const config = getLLMConfig();
+    // Guard on the exact transform: the same dims value at build and query time
+    // guarantees reduceDims produces matching-length vectors on both sides.
     if (
       parsed.provider !== config.provider ||
-      parsed.embeddingModel !== config.embeddingModel
+      parsed.embeddingModel !== config.embeddingModel ||
+      parsed.dims !== config.embeddingDims ||
+      parsed.quantization !== "int8"
     ) {
       return null;
     }
-    return parsed.chunks;
+    const decoded: IndexedChunk[] = [];
+    let expectedLen = -1;
+    for (const c of parsed.chunks) {
+      const embedding = decodeInt8(c.embedding);
+      // Reject a corrupt index rather than scoring against ragged vectors.
+      if (expectedLen === -1) expectedLen = embedding.length;
+      else if (embedding.length !== expectedLen) return null;
+      decoded.push({
+        id: c.id,
+        text: c.text,
+        path: c.path,
+        uri: c.uri,
+        embedding,
+      });
+    }
+    return decoded;
   } catch {
     return null;
   }
@@ -93,6 +147,7 @@ async function buildDocsIndex(force = false): Promise<void> {
   if (docsIndex.ready && !force) return;
 
   docsIndex.building = true;
+  indexUnavailableReason = null;
   try {
     // Prefer embeddings precomputed at build time - avoids re-embedding the
     // entire doc set on every cold start (the main cause of slow first chats).
@@ -113,24 +168,61 @@ async function buildDocsIndex(force = false): Promise<void> {
       return;
     }
 
+    // Reaching here means no usable prebuilt index was loaded (missing, empty,
+    // or built with a different provider/model/dims). Embedding the whole set on
+    // demand only completes in time for small doc sets; for large ones it runs
+    // one round trip per BATCH_SIZE chunks and blows past the serverless function
+    // limit, so the chat connects and then hangs forever with no answer. In
+    // production, fail fast with an actionable error instead of hanging (this
+    // also stops a client-forced \`refresh\` from burning embedding quota).
+    if (
+      process.env.NODE_ENV === "production" &&
+      chunks.length > RUNTIME_EMBED_MAX_CHUNKS
+    ) {
+      indexUnavailableReason =
+        \`Prebuilt embeddings index missing; refusing to embed \${chunks.length} \` +
+        \`chunks at request time (limit \${RUNTIME_EMBED_MAX_CHUNKS}). Run the build \` +
+        \`with an embedding API key set so scripts/build-docs-index.mts writes \` +
+        \`services/mcp/docs-index.json, then redeploy.\`;
+      console.error(\`[doccupine] \${indexUnavailableReason}\`);
+      throw new IndexNotBuiltError(
+        "The AI assistant is temporarily unavailable: its documentation search " +
+          "index has not been built for this deployment. If you are the site " +
+          "owner, redeploy with an embedding API key available at build time.",
+      );
+    }
+
+    // Small enough (or local dev): embed on demand. This still re-embeds on every
+    // cold start, so a prebuilt index is strongly preferred in production.
+    console.warn(
+      \`[doccupine] No prebuilt embeddings index found; embedding \${chunks.length} \` +
+        \`chunks on demand. Precompute at build time to avoid this on each cold start.\`,
+    );
+
     const config = getLLMConfig();
     const embeddings = createEmbeddings(config);
 
-    // Process embeddings in small batches to avoid exceeding token limits
+    // Process embeddings in small batches to avoid exceeding token limits.
+    // Reduce + quantize to the same int8 representation as the precomputed
+    // index so searchDocs scores identically on either path.
     const BATCH_SIZE = 10;
     const texts = chunks.map((c) => c.text);
-    const vectors: number[][] = [];
+    const built: IndexedChunk[] = [];
 
     for (let i = 0; i < texts.length; i += BATCH_SIZE) {
       const batch = texts.slice(i, i + BATCH_SIZE);
       const batchVectors = await embeddings.embedDocuments(batch);
-      vectors.push(...batchVectors);
+      for (let j = 0; j < batchVectors.length; j++) {
+        built.push({
+          ...chunks[i + j],
+          embedding: quantizeInt8(
+            reduceDims(batchVectors[j], config.embeddingDims),
+          ),
+        });
+      }
     }
 
-    docsIndex.chunks = chunks.map((c, i) => ({
-      ...c,
-      embedding: vectors[i],
-    }));
+    docsIndex.chunks = built;
     docsIndex.ready = true;
   } catch (error) {
     // Reset so the next call to ensureDocsIndex retries
@@ -164,7 +256,12 @@ export async function ensureDocsIndex(force = false): Promise<void> {
 
 // Eagerly start building the index on server startup if LLM is configured
 if (isLLMAvailable()) {
-  indexReady = buildDocsIndex();
+  const initialBuild = buildDocsIndex();
+  indexReady = initialBuild;
+  // A failed eager build (e.g. a missing prebuilt index on a large doc set) must
+  // not surface as an unhandledRejection at startup; the first request retries
+  // via ensureDocsIndex and returns a real error to the client.
+  void initialBuild.catch(() => {});
 }
 
 /** Cached embeddings instance for search queries */
@@ -186,12 +283,16 @@ export async function searchDocs(
 ): Promise<{ chunk: DocsChunk; score: number }[]> {
   await ensureDocsIndex();
 
-  const queryVector = await getEmbeddings().embedQuery(query);
+  // Reduce the query vector with the exact same transform used at index time so
+  // dimensions line up. Scoring runs directly against the int8 vectors - cosine
+  // is scale-invariant, so no dequantization is needed.
+  const rawQueryVector = await getEmbeddings().embedQuery(query);
+  const queryVector = reduceDims(rawQueryVector, getLLMConfig().embeddingDims);
 
   const scored = docsIndex.chunks
     .map((c) => ({
       chunk: { id: c.id, text: c.text, path: c.path, uri: c.uri },
-      score: cosineSim(queryVector, c.embedding),
+      score: cosineFloatInt8(queryVector, c.embedding),
     }))
     .sort((a, b) => b.score - a.score)
     .slice(0, limit);
@@ -202,10 +303,15 @@ export async function searchDocs(
 /**
  * Get the current index status
  */
-export function getIndexStatus(): { ready: boolean; chunkCount: number } {
+export function getIndexStatus(): {
+  ready: boolean;
+  chunkCount: number;
+  reason: string | null;
+} {
   return {
     ready: docsIndex.ready,
     chunkCount: docsIndex.chunks.length,
+    reason: indexUnavailableReason,
   };
 }
 

package/dist/templates/services/mcp/vector.d.ts

@@ -0,0 +1 @@
+export declare const vectorHelpersTemplate = "/**\n * Shared vector helpers for the docs embedding index.\n *\n * Both the build-time indexer (scripts/build-docs-index.mts) and the runtime\n * search path (services/mcp/server.ts) import these so query vectors and stored\n * vectors get the EXACT same transform - any mismatch would silently corrupt\n * ranking. Vectors are Matryoshka-truncated to a smaller dimension and stored\n * as int8, which shrinks each vector ~40x (and the overall index ~20x once\n * chunk text is counted) versus raw JSON floats. Large doc sets used to produce\n * 100MB+ indexes that OOM'd or timed out serverless cold starts (the chat would\n * connect, then idle forever).\n */\n\n/** L2-normalize a vector in place and return it. */\nexport function l2normalize(vector: number[]): number[] {\n  let norm = 0;\n  for (let i = 0; i < vector.length; i++) norm += vector[i] * vector[i];\n  norm = Math.sqrt(norm);\n  if (norm === 0) return vector;\n  for (let i = 0; i < vector.length; i++) vector[i] = vector[i] / norm;\n  return vector;\n}\n\n/**\n * Matryoshka dimension reduction: keep the first `dims` components and\n * renormalize. text-embedding-3-* and gemini-embedding-001 are MRL-trained, so\n * a renormalized prefix is a valid lower-dimension embedding. When `dims` is\n * <= 0 or >= the vector length, the vector is returned normalized at full\n * length. Operates on a copy so the caller's vector is left untouched.\n */\nexport function reduceDims(vector: number[], dims: number): number[] {\n  const target = dims > 0 && dims < vector.length ? dims : vector.length;\n  return l2normalize(vector.slice(0, target));\n}\n\n/**\n * Quantize a float vector to int8 using a per-vector max-abs scale. Cosine\n * similarity is scale-invariant, so the scale factor never has to be stored:\n * cosine(query, scale * q) === cosine(query, q). Callers score directly against\n * the int8 array via cosineFloatInt8, so there is no dequantization step.\n */\nexport function quantizeInt8(vector: number[]): Int8Array {\n  let maxAbs = 0;\n  for (let i = 0; i < vector.length; i++) {\n    const a = Math.abs(vector[i]);\n    if (a > maxAbs) maxAbs = a;\n  }\n  const scale = maxAbs > 0 ? 127 / maxAbs : 0;\n  const q = new Int8Array(vector.length);\n  for (let i = 0; i < vector.length; i++) {\n    // Clamp before assignment: Int8Array wraps out-of-range values (200 -> -56)\n    // instead of clamping, which would flip a component's sign.\n    let s = Math.round(vector[i] * scale);\n    if (s > 127) s = 127;\n    else if (s < -127) s = -127;\n    q[i] = s;\n  }\n  return q;\n}\n\n/** Encode an int8 vector as base64 for compact JSON storage. */\nexport function encodeInt8(q: Int8Array): string {\n  return Buffer.from(q.buffer, q.byteOffset, q.byteLength).toString(\"base64\");\n}\n\n/** Decode a base64 int8 vector produced by encodeInt8. */\nexport function decodeInt8(b64: string): Int8Array {\n  const buf = Buffer.from(b64, \"base64\");\n  return new Int8Array(buf.buffer, buf.byteOffset, buf.byteLength);\n}\n\n/**\n * Cosine similarity between a float query vector and an int8 stored vector.\n * Both must share the same length. The int8 scale cancels out in the\n * normalization, so scoring against the raw int8 array is exact up to the\n * quantization rounding error (negligible for top-K retrieval).\n */\nexport function cosineFloatInt8(query: number[], stored: Int8Array): number {\n  const len = Math.min(query.length, stored.length);\n  let dot = 0;\n  let na = 0;\n  let nb = 0;\n  for (let i = 0; i < len; i++) {\n    const x = query[i];\n    const y = stored[i];\n    dot += x * y;\n    na += x * x;\n    nb += y * y;\n  }\n  if (na === 0 || nb === 0) return 0;\n  return dot / (Math.sqrt(na) * Math.sqrt(nb));\n}\n";

package/dist/templates/services/mcp/vector.js

@@ -0,0 +1,93 @@
+export const vectorHelpersTemplate = `/**
+ * Shared vector helpers for the docs embedding index.
+ *
+ * Both the build-time indexer (scripts/build-docs-index.mts) and the runtime
+ * search path (services/mcp/server.ts) import these so query vectors and stored
+ * vectors get the EXACT same transform - any mismatch would silently corrupt
+ * ranking. Vectors are Matryoshka-truncated to a smaller dimension and stored
+ * as int8, which shrinks each vector ~40x (and the overall index ~20x once
+ * chunk text is counted) versus raw JSON floats. Large doc sets used to produce
+ * 100MB+ indexes that OOM'd or timed out serverless cold starts (the chat would
+ * connect, then idle forever).
+ */
+
+/** L2-normalize a vector in place and return it. */
+export function l2normalize(vector: number[]): number[] {
+  let norm = 0;
+  for (let i = 0; i < vector.length; i++) norm += vector[i] * vector[i];
+  norm = Math.sqrt(norm);
+  if (norm === 0) return vector;
+  for (let i = 0; i < vector.length; i++) vector[i] = vector[i] / norm;
+  return vector;
+}
+
+/**
+ * Matryoshka dimension reduction: keep the first \`dims\` components and
+ * renormalize. text-embedding-3-* and gemini-embedding-001 are MRL-trained, so
+ * a renormalized prefix is a valid lower-dimension embedding. When \`dims\` is
+ * <= 0 or >= the vector length, the vector is returned normalized at full
+ * length. Operates on a copy so the caller's vector is left untouched.
+ */
+export function reduceDims(vector: number[], dims: number): number[] {
+  const target = dims > 0 && dims < vector.length ? dims : vector.length;
+  return l2normalize(vector.slice(0, target));
+}
+
+/**
+ * Quantize a float vector to int8 using a per-vector max-abs scale. Cosine
+ * similarity is scale-invariant, so the scale factor never has to be stored:
+ * cosine(query, scale * q) === cosine(query, q). Callers score directly against
+ * the int8 array via cosineFloatInt8, so there is no dequantization step.
+ */
+export function quantizeInt8(vector: number[]): Int8Array {
+  let maxAbs = 0;
+  for (let i = 0; i < vector.length; i++) {
+    const a = Math.abs(vector[i]);
+    if (a > maxAbs) maxAbs = a;
+  }
+  const scale = maxAbs > 0 ? 127 / maxAbs : 0;
+  const q = new Int8Array(vector.length);
+  for (let i = 0; i < vector.length; i++) {
+    // Clamp before assignment: Int8Array wraps out-of-range values (200 -> -56)
+    // instead of clamping, which would flip a component's sign.
+    let s = Math.round(vector[i] * scale);
+    if (s > 127) s = 127;
+    else if (s < -127) s = -127;
+    q[i] = s;
+  }
+  return q;
+}
+
+/** Encode an int8 vector as base64 for compact JSON storage. */
+export function encodeInt8(q: Int8Array): string {
+  return Buffer.from(q.buffer, q.byteOffset, q.byteLength).toString("base64");
+}
+
+/** Decode a base64 int8 vector produced by encodeInt8. */
+export function decodeInt8(b64: string): Int8Array {
+  const buf = Buffer.from(b64, "base64");
+  return new Int8Array(buf.buffer, buf.byteOffset, buf.byteLength);
+}
+
+/**
+ * Cosine similarity between a float query vector and an int8 stored vector.
+ * Both must share the same length. The int8 scale cancels out in the
+ * normalization, so scoring against the raw int8 array is exact up to the
+ * quantization rounding error (negligible for top-K retrieval).
+ */
+export function cosineFloatInt8(query: number[], stored: Int8Array): number {
+  const len = Math.min(query.length, stored.length);
+  let dot = 0;
+  let na = 0;
+  let nb = 0;
+  for (let i = 0; i < len; i++) {
+    const x = query[i];
+    const y = stored[i];
+    dot += x * y;
+    na += x * x;
+    nb += y * y;
+  }
+  if (na === 0 || nb === 0) return 0;
+  return dot / (Math.sqrt(na) * Math.sqrt(nb));
+}
+`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "doccupine",
-  "version": "0.0.119",
+  "version": "0.0.121",
   "description": "Free and open-source documentation platform. Write MDX, get a production-ready site with AI chat, built-in components, and an MCP server - in one command.",
   "main": "dist/index.js",
   "bin": {