llm-kb 0.4.2 → 0.6.0

package/README.md

@@ -15,20 +15,25 @@ That's it. PDFs get parsed, an index is built, and an interactive chat opens —
 
 ## Authentication
 
-Two options (you need one):
+Three options (you need one):
 
-**Option 1 — Pi SDK (recommended)**
+**Option 1 — OpenRouter API key (recommended)**
+```bash
+export OPENROUTER_API_KEY=sk-or-...
+```
+
+**Option 2 — Pi SDK**
 ```bash
 npm install -g @mariozechner/pi-coding-agent
 pi   # run once to authenticate
 ```
 
-**Option 2 — Anthropic API key**
+**Option 3 — Anthropic API key**
 ```bash
 export ANTHROPIC_API_KEY=sk-ant-...
 ```
 
-If neither is configured, `llm-kb` shows a clear error with setup instructions.
+If none is configured, `llm-kb` shows a clear error with setup instructions.
 
 ## What It Does
 
@@ -39,7 +44,7 @@ llm-kb run ./my-documents
 ```
 
 ```
-llm-kb v0.4.1
+llm-kb v0.5.0
 
 Scanning ./my-documents...
   Found 9 files (9 PDF)
@@ -156,7 +161,7 @@ Knowledge Base Status
   Articles: 15 compiled
   Outputs: 2 saved answers
   Models:  claude-sonnet-4-6 (query)  claude-haiku-4-5 (index)
-  Auth:    Pi SDK
+  Auth:    OpenRouter
 ```
 
 ## The Three-Layer Architecture

package/bin/{chunk-DHOXVEIR.js → chunk-3WBSKCCH.js}

@@ -6776,16 +6776,6 @@ var require_dist = __commonJS({
   }
 });
 
-// src/indexer.ts
-import {
-  createAgentSession,
-  createBashTool,
-  createReadTool,
-  createWriteTool,
-  DefaultResourceLoader,
-  SettingsManager
-} from "@mariozechner/pi-coding-agent";
-
 // node_modules/@mariozechner/pi-ai/dist/api-registry.js
 var apiProviderRegistry = /* @__PURE__ */ new Map();
 function wrapStream(api, stream) {
@@ -7112,8 +7102,99 @@ if (canUseRuntimeCodegen()) {
   }
 }
 
-// src/indexer.ts
-import { readdir, readFile } from "fs/promises";
+// src/model-resolver.ts
+import { AuthStorage } from "@mariozechner/pi-coding-agent";
+var ANTHROPIC_TO_OPENROUTER = {
+  "claude-haiku-4-5": "anthropic/claude-haiku-4.5",
+  "claude-sonnet-4-6": "anthropic/claude-sonnet-4.6",
+  "claude-sonnet-4-5": "anthropic/claude-sonnet-4.5",
+  "claude-sonnet-4-0": "anthropic/claude-sonnet-4",
+  "claude-opus-4-5": "anthropic/claude-opus-4.5"
+};
+var OPENROUTER_TO_ANTHROPIC = Object.fromEntries(
+  Object.entries(ANTHROPIC_TO_OPENROUTER).map(([anthropic, openrouter]) => [openrouter, anthropic])
+);
+var PURPOSE_FALLBACKS = {
+  index: ["claude-haiku-4-5", "anthropic/claude-haiku-4.5", "claude-sonnet-4-6", "anthropic/claude-sonnet-4.6"],
+  query: ["claude-sonnet-4-6", "anthropic/claude-sonnet-4.6", "claude-sonnet-4-5", "anthropic/claude-sonnet-4.5"],
+  wiki: ["claude-haiku-4-5", "anthropic/claude-haiku-4.5", "claude-sonnet-4-6", "anthropic/claude-sonnet-4.6"],
+  eval: ["claude-haiku-4-5", "anthropic/claude-haiku-4.5", "claude-sonnet-4-6", "anthropic/claude-sonnet-4.6"],
+  generic: ["claude-haiku-4-5", "anthropic/claude-haiku-4.5", "claude-sonnet-4-6", "anthropic/claude-sonnet-4.6"]
+};
+function stripOpenAI(prefixOrId) {
+  return prefixOrId.replace(/^openai\//, "");
+}
+function modelCandidates(modelId, purpose) {
+  const ids = /* @__PURE__ */ new Set();
+  ids.add(modelId);
+  const mappedOpenRouter = ANTHROPIC_TO_OPENROUTER[modelId];
+  if (mappedOpenRouter) ids.add(mappedOpenRouter);
+  const mappedAnthropic = OPENROUTER_TO_ANTHROPIC[modelId];
+  if (mappedAnthropic) ids.add(mappedAnthropic);
+  if (modelId.startsWith("openai/")) ids.add(stripOpenAI(modelId));
+  else ids.add(`openai/${modelId}`);
+  for (const fallback of PURPOSE_FALLBACKS[purpose]) ids.add(fallback);
+  return [...ids];
+}
+function providerOrder(modelId) {
+  if (modelId.startsWith("openai/") || modelId.startsWith("gpt-")) {
+    return ["openrouter", "openai", "anthropic"];
+  }
+  if (modelId.startsWith("anthropic/") || modelId.startsWith("claude-")) {
+    return ["openrouter", "anthropic", "openai"];
+  }
+  return ["openrouter", "openai", "anthropic"];
+}
+function resolveIdForProvider(provider, candidateId) {
+  switch (provider) {
+    case "anthropic": {
+      const ids = [candidateId];
+      const mapped = OPENROUTER_TO_ANTHROPIC[candidateId];
+      if (mapped) ids.push(mapped);
+      return [...new Set(ids.filter((id) => !id.startsWith("openai/")))];
+    }
+    case "openrouter": {
+      const ids = [candidateId];
+      const mapped = ANTHROPIC_TO_OPENROUTER[candidateId];
+      if (mapped) ids.unshift(mapped);
+      if (candidateId.startsWith("gpt-")) ids.unshift(`openai/${candidateId}`);
+      return [...new Set(ids)];
+    }
+    case "openai": {
+      return [stripOpenAI(candidateId)].filter((id) => !id.startsWith("claude-") && !id.startsWith("anthropic/"));
+    }
+  }
+}
+async function findModelForProvider(provider, candidateId, storage) {
+  const key = await storage.getApiKey(provider);
+  if (!key) return void 0;
+  const available = getModels(provider);
+  for (const id of resolveIdForProvider(provider, candidateId)) {
+    const model = available.find((m) => m.id === id);
+    if (model) return model;
+  }
+  return void 0;
+}
+async function resolveModelCandidates(modelId, authStorage, purpose = "generic") {
+  const storage = authStorage ?? AuthStorage.create();
+  const resolved = [];
+  const seen = /* @__PURE__ */ new Set();
+  for (const candidateId of modelCandidates(modelId, purpose)) {
+    for (const provider of providerOrder(candidateId)) {
+      const model = await findModelForProvider(provider, candidateId, storage);
+      if (!model) continue;
+      const key = `${provider}:${model.id}`;
+      if (seen.has(key)) continue;
+      seen.add(key);
+      resolved.push({ provider, candidateId, model });
+    }
+  }
+  return resolved;
+}
+async function getApiKeyForProvider(provider, authStorage) {
+  const storage = authStorage ?? AuthStorage.create();
+  return storage.getApiKey(provider);
+}
 
 // src/session-store.ts
 import { SessionManager } from "@mariozechner/pi-coding-agent";
@@ -7147,115 +7228,11 @@ function getNodeModulesPath() {
   return join2(process.cwd(), "node_modules");
 }
 
-// src/indexer.ts
-import { join as join3 } from "path";
-function buildAgentsContent(sourcesDir, files) {
-  const sourceList = files.filter((f) => f.endsWith(".md")).map((f) => `  - ${f}`).join("\n");
-  return `# llm-kb Knowledge Base
-
-## How to access documents
-
-### PDFs (pre-parsed)
-PDFs have been parsed to markdown with bounding boxes.
-Read the markdown versions in \`.llm-kb/wiki/sources/\` instead of the raw PDFs.
-
-Available parsed sources:
-${sourceList}
-
-### Other file types (Excel, Word, PowerPoint)
-You have bash and read tools. Use bash to run Node.js scripts.
-Libraries are pre-installed via require().
-
-For .docx (structured XML \u2014 ZIP containing word/document.xml):
-  const AdmZip = require('adm-zip');
-  const zip = new AdmZip('file.docx');
-  const xml = zip.readAsText('word/document.xml');
-  // Parse XML to extract headings and first paragraphs for summary
-
-For .xlsx use exceljs:
-  const ExcelJS = require('exceljs');
-  const wb = new ExcelJS.Workbook();
-  await wb.xlsx.readFile('file.xlsx');
-  const sheet = wb.getWorksheet(1);
-
-For .pptx use officeparser:
-  const officeparser = require('officeparser');
-  const text = await officeparser.parseOfficeAsync('file.pptx');
-
-## Index file
-Write the index to \`.llm-kb/wiki/index.md\`.
-
-The index should be a markdown file with:
-1. A title and last-updated timestamp
-2. A summary table with columns: Source, Type, Pages/Size, Summary, Key Topics
-3. Each source gets a one-line summary (read the first ~500 chars of each file to generate it)
-4. Total word count across all sources
-`;
-}
-async function buildIndex(folder, sourcesDir, onOutput, authStorage, modelId) {
-  const files = await readdir(sourcesDir);
-  const mdFiles = files.filter((f) => f.endsWith(".md"));
-  if (mdFiles.length === 0) {
-    throw new Error("No source files found to index");
-  }
-  const agentsContent = buildAgentsContent(sourcesDir, files);
-  const nodeModulesPath = getNodeModulesPath();
-  process.env.NODE_PATH = nodeModulesPath;
-  const loader = new DefaultResourceLoader({
-    cwd: folder,
-    agentsFilesOverride: (current) => ({
-      agentsFiles: [
-        ...current.agentsFiles,
-        { path: ".llm-kb/AGENTS.md", content: agentsContent }
-      ]
-    })
-  });
-  await loader.reload();
-  const model = modelId ? getModels("anthropic").find((m) => m.id === modelId) : void 0;
-  const { session } = await createAgentSession({
-    cwd: folder,
-    resourceLoader: loader,
-    tools: [
-      createReadTool(folder),
-      createBashTool(folder),
-      createWriteTool(folder)
-    ],
-    sessionManager: await createKBSession(folder),
-    settingsManager: SettingsManager.inMemory({
-      compaction: { enabled: false }
-    }),
-    ...authStorage ? { authStorage } : {},
-    ...model ? { model } : {}
-  });
-  if (onOutput) {
-    session.subscribe((event) => {
-      if (event.type === "message_update" && event.assistantMessageEvent.type === "text_delta") {
-        onOutput(event.assistantMessageEvent.delta);
-      }
-    });
-  }
-  session.setSessionName(`index: ${(/* @__PURE__ */ new Date()).toISOString()}`);
-  const prompt = `Read each file in .llm-kb/wiki/sources/ (one at a time, just the first 500 characters of each).
-Then write .llm-kb/wiki/index.md with a summary table of all sources.
-
-Include: Source filename, Type (PDF/Excel/Word/etc), Pages (from the JSON if available), a one-line summary, and key topics.
-Add a total word count estimate at the bottom.`;
-  await session.prompt(prompt);
-  const indexPath = join3(sourcesDir, "..", "index.md");
-  try {
-    const content = await readFile(indexPath, "utf-8");
-    session.dispose();
-    return content;
-  } catch {
-    session.dispose();
-    throw new Error("Agent did not create index.md");
-  }
-}
-
 export {
   completeSimple,
+  resolveModelCandidates,
+  getApiKeyForProvider,
   continueKBSession,
   createKBSession,
-  getNodeModulesPath,
-  buildIndex
+  getNodeModulesPath
 };

package/bin/chunk-EZ7LPPEP.js

@@ -0,0 +1,218 @@
+import {
+  createKBSession,
+  getNodeModulesPath,
+  resolveModelCandidates
+} from "./chunk-3WBSKCCH.js";
+
+// src/indexer.ts
+import {
+  createAgentSession,
+  createBashTool,
+  createReadTool,
+  createWriteTool,
+  DefaultResourceLoader,
+  SettingsManager
+} from "@mariozechner/pi-coding-agent";
+import { readdir, readFile } from "fs/promises";
+import { join } from "path";
+function buildAgentsContent(sourcesDir, files) {
+  const sourceList = files.filter((f) => f.endsWith(".md")).map((f) => `  - ${f}`).join("\n");
+  return `# llm-kb Knowledge Base
+
+## How to access documents
+
+### PDFs (pre-parsed)
+PDFs have been parsed to markdown with bounding boxes.
+Read the markdown versions in \`.llm-kb/wiki/sources/\` instead of the raw PDFs.
+
+Available parsed sources:
+${sourceList}
+
+### Other file types (Excel, Word, PowerPoint)
+You have bash and read tools. Use bash to run Node.js scripts.
+Libraries are pre-installed via require().
+
+For .docx (structured XML \u2014 ZIP containing word/document.xml):
+  const AdmZip = require('adm-zip');
+  const zip = new AdmZip('file.docx');
+  const xml = zip.readAsText('word/document.xml');
+  // Parse XML to extract headings and first paragraphs for summary
+
+For .xlsx use exceljs:
+  const ExcelJS = require('exceljs');
+  const wb = new ExcelJS.Workbook();
+  await wb.xlsx.readFile('file.xlsx');
+  const sheet = wb.getWorksheet(1);
+
+For .pptx use officeparser:
+  const officeparser = require('officeparser');
+  const text = await officeparser.parseOfficeAsync('file.pptx');
+
+## Index file
+Write the index to \`.llm-kb/wiki/index.md\`.
+
+The index should be a markdown file with:
+1. A title and last-updated timestamp
+2. A summary table with columns: Source, Type, Pages/Size, Summary, Key Topics
+3. Each source gets a one-line summary (read the first ~500 chars of each file to generate it)
+4. Total word count across all sources
+`;
+}
+async function buildIndex(folder, sourcesDir, onOutput, authStorage, modelId) {
+  const files = await readdir(sourcesDir);
+  const mdFiles = files.filter((f) => f.endsWith(".md"));
+  const jsonFiles = files.filter((f) => f.endsWith(".json") && !f.endsWith(".pages"));
+  if (mdFiles.length === 0) {
+    throw new Error("No source files found to index");
+  }
+  const snippets = [];
+  const total = mdFiles.length;
+  const cols = process.stdout.columns || 80;
+  for (let i = 0; i < mdFiles.length; i++) {
+    const f = mdFiles[i];
+    const pct = Math.round((i + 1) / total * 100);
+    const name = f.length > 30 ? f.slice(0, 27) + "..." : f;
+    process.stdout.write(`\r  Reading sources... ${i + 1}/${total} (${pct}%) ${name}`.padEnd(cols));
+    try {
+      const content = await readFile(join(sourcesDir, f), "utf-8");
+      const preview = content.slice(0, 800);
+      const jsonName = f.replace(/\.md$/, ".json");
+      let pages = 0;
+      if (jsonFiles.includes(jsonName)) {
+        try {
+          const jsonHead = await readFile(join(sourcesDir, jsonName), "utf-8");
+          const match = jsonHead.match(/"totalPages"\s*:\s*(\d+)/);
+          if (match) pages = parseInt(match[1], 10);
+        } catch {
+        }
+      }
+      snippets.push(`### ${f}${pages > 0 ? ` (${pages} pages)` : ""}
+${preview}
+`);
+    } catch {
+      snippets.push(`### ${f}
+(could not read)
+`);
+    }
+  }
+  process.stdout.write(`\r${"".padEnd(cols)}\r`);
+  process.stdout.write(`  Read ${mdFiles.length} source previews
+`);
+  const BATCH_SIZE = 100;
+  const batches = [];
+  for (let i = 0; i < snippets.length; i += BATCH_SIZE) {
+    batches.push(snippets.slice(i, i + BATCH_SIZE));
+  }
+  const agentsContent = buildAgentsContent(sourcesDir, files);
+  const nodeModulesPath = getNodeModulesPath();
+  process.env.NODE_PATH = nodeModulesPath;
+  const candidates = modelId ? await resolveModelCandidates(modelId, authStorage, "index") : [];
+  if (modelId && candidates.length === 0) {
+    throw new Error(`No usable model found for '${modelId}'. Configure Anthropic, OpenRouter, or OpenAI credentials.`);
+  }
+  const indexPath = join(sourcesDir, "..", "index.md");
+  const attemptCandidates = candidates.length > 0 ? candidates : [{ provider: "default", candidateId: "default", model: void 0 }];
+  const batchResults = [];
+  for (let b = 0; b < batches.length; b++) {
+    const batch = batches[b];
+    const batchLabel = batches.length > 1 ? ` (batch ${b + 1}/${batches.length})` : "";
+    process.stdout.write(`  Generating index${batchLabel}...
+`);
+    const batchContent = batch.join("\n---\n\n");
+    const prompt = batches.length === 1 ? `Here are previews of all ${mdFiles.length} source files in this knowledge base. Generate a summary table in markdown.
+
+${batchContent}
+
+Write .llm-kb/wiki/index.md with:
+1. Title and last-updated timestamp
+2. A markdown table with columns: Source, Type, Pages, Summary, Key Topics
+3. One row per source with a one-line summary
+4. Total count at the bottom
+
+Do NOT read any files \u2014 all the data you need is above.` : b < batches.length - 1 ? `Here are previews of source files ${b * BATCH_SIZE + 1}-${Math.min((b + 1) * BATCH_SIZE, mdFiles.length)} of ${mdFiles.length}. Generate summary table rows ONLY (no header, no footer).
+
+${batchContent}
+
+Output ONLY markdown table rows \u2014 one per source. Columns: Source, Type, Pages, Summary, Key Topics.
+Do NOT read any files.` : `Here are the remaining source file previews (${b * BATCH_SIZE + 1}-${mdFiles.length} of ${mdFiles.length}).
+
+${batchContent}
+
+Output ONLY markdown table rows for these sources. Columns: Source, Type, Pages, Summary, Key Topics.
+
+Then combine with the previous batch results below and write the final .llm-kb/wiki/index.md:
+
+Previous batch rows:
+${batchResults.join("\n")}
+
+Write the complete index.md with title, timestamp, full table (header + all rows), and total count.`;
+    let lastError;
+    for (let i = 0; i < attemptCandidates.length; i++) {
+      const candidate = attemptCandidates[i];
+      const loader = new DefaultResourceLoader({
+        cwd: folder,
+        agentsFilesOverride: (current) => ({
+          agentsFiles: [
+            ...current.agentsFiles,
+            { path: ".llm-kb/AGENTS.md", content: agentsContent }
+          ]
+        })
+      });
+      await loader.reload();
+      const { session } = await createAgentSession({
+        cwd: folder,
+        resourceLoader: loader,
+        tools: [
+          createReadTool(folder),
+          createBashTool(folder),
+          createWriteTool(folder)
+        ],
+        sessionManager: await createKBSession(folder),
+        settingsManager: SettingsManager.inMemory({
+          compaction: { enabled: false }
+        }),
+        ...authStorage ? { authStorage } : {},
+        ...candidate.model ? { model: candidate.model } : {}
+      });
+      if (onOutput) {
+        session.subscribe((event) => {
+          if (event.type === "message_update" && event.assistantMessageEvent.type === "text_delta") {
+            onOutput(event.assistantMessageEvent.delta);
+          }
+        });
+      }
+      session.setSessionName(`index: ${(/* @__PURE__ */ new Date()).toISOString()}`);
+      try {
+        await session.prompt(prompt);
+        if (batches.length === 1 || b === batches.length - 1) {
+          const content = await readFile(indexPath, "utf-8");
+          session.dispose();
+          return content;
+        } else {
+          const messages = session.state.messages;
+          const lastAssistant = [...messages].reverse().find((m) => m.role === "assistant");
+          const text = lastAssistant?.content?.filter((b2) => b2.type === "text").map((b2) => b2.text).join("") ?? "";
+          batchResults.push(text);
+          session.dispose();
+          break;
+        }
+      } catch (error) {
+        lastError = error;
+        session.dispose();
+        const next = attemptCandidates[i + 1];
+        if (next) {
+          const detail = error instanceof Error ? error.message : String(error);
+          console.warn(`  Index attempt failed on ${candidate.provider}:${candidate.model?.id ?? candidate.candidateId} (${detail}). Retrying with ${next.provider}:${next.model?.id ?? next.candidateId}...`);
+          continue;
+        }
+        if (lastError instanceof Error) throw lastError;
+        throw new Error("Agent did not create index.md");
+      }
+    }
+  }
+  throw new Error("Agent did not create index.md");
+}
+
+export {
+  buildIndex
+};