@every-app/sdk 0.3.0-rc.1 → 0.3.0-rc.3

package/dist/internal/catalog/inference.js

@@ -0,0 +1,457 @@
+export const INFERENCE = {
+    // Bearer is the documented header today; x-api-key is the legacy fallback. A workspace-scoped key keeps anthropic-workspace-id harmless.
+    anthropic: {
+        mode: "ai",
+        label: "Anthropic",
+        description: "Claude models through the Messages API.",
+        auth: {
+            kind: "bearer",
+            secret: "ANTHROPIC_API_KEY",
+        },
+        docs: "https://platform.claude.com/settings/keys",
+        api: "https://platform.claude.com/docs/en/api/overview",
+        origin: "https://api.anthropic.com",
+        inferenceRoutes: [
+            "/v1/messages",
+            "/v1/messages/count_tokens",
+            // Anthropic's OpenAI-compatible chat completions (its docs: not a
+            // long-term solution; `strict` is ignored): what the default route
+            // maps onto when Anthropic is the Gateway's default.
+            "/v1/chat/completions",
+        ],
+        unstoredRoutes: ["/v1/chat/completions"],
+        // The Messages API is the shape an app pins `anthropic` for.
+        shape: "anthropic",
+        chat: { path: "/v1/chat/completions", model: "claude-haiku-5-5" },
+        aig: { slug: "anthropic" },
+        hint: "An API key from the Claude Console's API keys page; the Gateway sends it as a bearer token. Apps send the anthropic-version header themselves.",
+        secrets: [
+            {
+                name: "ANTHROPIC_API_KEY",
+                hint: "A key scoped to one workspace, so no call needs an anthropic-workspace-id header.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/v1/models",
+            expect: [200],
+            free: true,
+            note: "Send an anthropic-version header, as every call does.",
+        },
+        rate: {
+            note: "Requests and tokens per minute by usage tier, on the Console's Rate limits page; token counting is free, with its own per-minute limit.",
+        },
+        example: "summarize this support thread in three bullet points",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "llm"],
+    },
+    gemini: {
+        mode: "ai",
+        label: "Google Gemini",
+        description: "Google's Gemini models through their OpenAI-compatible endpoint: chat, embeddings, images.",
+        auth: {
+            kind: "bearer",
+            secret: "GEMINI_API_KEY",
+        },
+        docs: "https://aistudio.google.com/apikey",
+        api: "https://ai.google.dev/gemini-api/docs/openai",
+        origin: "https://generativelanguage.googleapis.com",
+        inferenceRoutes: [
+            "/v1beta/openai/chat/completions",
+            "/v1beta/openai/embeddings",
+            "/v1beta/openai/images/generations",
+        ],
+        chat: {
+            path: "/v1beta/openai/chat/completions",
+            model: "gemini-3.1-flash-lite",
+        },
+        embeddings: {
+            path: "/v1beta/openai/embeddings",
+            model: "gemini-embedding-001",
+        },
+        // No `aig`: AI Gateway documents Google AI Studio's native form only
+        // (/v1/models/{model}:generateContent), not these OpenAI-shaped paths.
+        hint: "An API key from Google AI Studio; the Gateway sends it as a bearer token. Apps call Gemini's OpenAI-shaped paths under /v1beta/openai.",
+        secrets: [
+            {
+                name: "GEMINI_API_KEY",
+                hint: "A Google AI Studio key restricted to the Gemini API only.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/v1beta/openai/models",
+            expect: [200],
+            free: true,
+        },
+        rate: {
+            note: "Per project, not per key: requests and tokens per minute and requests per day, by tier (Free, then Tiers 1 to 3).",
+        },
+        example: "describe what is in this product photo",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "llm", "embeddings", "images"],
+    },
+    mistral: {
+        mode: "ai",
+        label: "Mistral AI",
+        description: "Mistral's models: chat, code completion, embeddings, moderation, OCR, speech.",
+        auth: {
+            kind: "bearer",
+            secret: "MISTRAL_API_KEY",
+        },
+        docs: "https://console.mistral.ai",
+        api: "https://docs.mistral.ai/api/",
+        origin: "https://api.mistral.ai",
+        inferenceRoutes: [
+            "/v1/chat/completions",
+            "/v1/fim/completions",
+            "/v1/embeddings",
+            "/v1/moderations",
+            "/v1/chat/moderations",
+            "/v1/classifications",
+            "/v1/chat/classifications",
+            "/v1/ocr",
+            "/v1/audio/transcriptions",
+            "/v1/audio/speech",
+        ],
+        chat: { path: "/v1/chat/completions", model: "mistral-small-latest" },
+        embeddings: { path: "/v1/embeddings", model: "mistral-embed" },
+        aig: { slug: "mistral" },
+        hint: "An API key from API Keys in the Mistral Studio console; the Gateway sends it as a bearer token. Free mode needs no card; usage and rate limits apply.",
+        secrets: [
+            {
+                name: "MISTRAL_API_KEY",
+                hint: "A key from API Keys in the Studio console, with an expiry date.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/v1/models",
+            expect: [200],
+            free: true,
+        },
+        example: "extract the text from this scanned invoice",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "llm", "ocr", "embeddings"],
+    },
+    groq: {
+        mode: "ai",
+        label: "Groq",
+        description: "Open models on Groq through OpenAI's API shape: chat, responses, transcription, speech.",
+        auth: {
+            kind: "bearer",
+            secret: "GROQ_API_KEY",
+        },
+        docs: "https://console.groq.com/keys",
+        api: "https://console.groq.com/docs/api-reference",
+        origin: "https://api.groq.com",
+        inferenceRoutes: [
+            "/openai/v1/chat/completions",
+            "/openai/v1/responses",
+            "/openai/v1/audio/transcriptions",
+            "/openai/v1/audio/translations",
+            "/openai/v1/audio/speech",
+        ],
+        chat: { path: "/openai/v1/chat/completions", model: "openai/gpt-oss-20b" },
+        // AI Gateway's Groq endpoint takes the path without `/openai/v1`.
+        aig: { slug: "groq", trim: "/openai/v1" },
+        hint: "An API key from the Groq console; the Gateway sends it as a bearer token. Apps call the OpenAI-shaped paths under /openai/v1.",
+        secrets: [
+            {
+                name: "GROQ_API_KEY",
+                hint: "A key from the API Keys page of the Groq console.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/openai/v1/models",
+            expect: [200],
+            free: true,
+        },
+        rate: {
+            note: "Requests and tokens per minute and per day, per model and plan, for the whole organization; the console Limits page shows them.",
+        },
+        example: "transcribe this voice memo",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "llm", "speech"],
+    },
+    cohere: {
+        mode: "ai",
+        label: "Cohere",
+        description: "Cohere's models: chat, embeddings, reranking, classification, transcription, document parsing.",
+        auth: {
+            kind: "bearer",
+            secret: "COHERE_API_KEY",
+        },
+        docs: "https://dashboard.cohere.com/api-keys",
+        api: "https://docs.cohere.com/reference/about",
+        origin: "https://api.cohere.com",
+        inferenceRoutes: [
+            "/v2/chat",
+            "/v2/embed",
+            "/v2/rerank",
+            "/v1/classify",
+            "/v2/audio/transcriptions",
+            "/v2/parse",
+            "/v1/tokenize",
+            "/v1/detokenize",
+        ],
+        // Cohere's own v2 shape: pinned by an app that speaks it, never the
+        // default (its OpenAI-compatible endpoint is on another host,
+        // api.cohere.ai). No `aig`: AI Gateway documents v1 paths only.
+        shape: "other",
+        hint: "An API key from the Cohere dashboard; the Gateway sends it as a bearer token. Trial keys are free and limited; production keys are paid.",
+        secrets: [
+            {
+                name: "COHERE_API_KEY",
+                hint: "A production key; a trial key is limited to 1,000 calls a month.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/v1/models",
+            expect: [200],
+            free: true,
+        },
+        rate: {
+            note: "Production: 500 calls a minute for Command A chat and most endpoints, rerank 1,000. Trial: 20 chat calls a minute, 1,000 calls a month.",
+        },
+        example: "rerank these search results for the query refund policy",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "llm", "embeddings", "rerank"],
+    },
+    voyage: {
+        mode: "api_key",
+        label: "Voyage AI",
+        description: "Embeddings and rerankers for search and retrieval: text, code, multimodal, contextualized chunks.",
+        auth: {
+            kind: "bearer",
+            secret: "VOYAGE_API_KEY",
+        },
+        docs: "https://dashboard.voyageai.com/organization/api-keys",
+        api: "https://docs.voyageai.com/reference/embeddings-api",
+        hosts: ["api.voyageai.com"],
+        data: "per-call",
+        access: ["read"],
+        readMethods: ["POST"],
+        accountWidePaths: [
+            ["v1", "files"],
+            ["v1", "batches"],
+        ],
+        hint: "A secret key from the Voyage dashboard; the Gateway sends it as a bearer token. Apps POST to /v1/embeddings, /v1/multimodalembeddings, /v1/contextualizedembeddings and /v1/rerank.",
+        secrets: [
+            {
+                name: "VOYAGE_API_KEY",
+                hint: "A secret key from API keys in the Voyage dashboard.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/v1/batches",
+            expect: [200],
+            free: true,
+            note: "Lists the organization's batches, so the Gateway refuses it for apps; run it from the person's terminal.",
+        },
+        cost: {
+            unit: "token",
+            note: "voyage-4 $0.06 per million tokens; the first 200 million tokens of the voyage-4 family are free per account.",
+        },
+        rate: {
+            perMinute: 2000,
+            note: "2,000 requests a minute at tier 1, doubled at tier 2, tripled at tier 3; tokens per minute vary by model.",
+        },
+        example: "embed these product descriptions for semantic search",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["embeddings", "rerank", "search"],
+    },
+    perplexity: {
+        mode: "api_key",
+        label: "Perplexity",
+        description: "Web-grounded answers from the Agent API, raw web search results, and embeddings.",
+        auth: {
+            kind: "bearer",
+            secret: "PERPLEXITY_API_KEY",
+        },
+        docs: "https://console.perplexity.ai/project/keys",
+        api: "https://docs.perplexity.ai/api-reference/agent-post",
+        hosts: ["api.perplexity.ai"],
+        data: "per-call",
+        access: ["read"],
+        readMethods: ["GET", "POST"],
+        accountWidePaths: [
+            ["v1", "batches"],
+            ["v1", "files"],
+            ["v1", "skills"],
+            ["v1", "analytics"],
+            ["v2", "analytics"],
+            ["generate_auth_token"],
+            ["revoke_auth_token"],
+        ],
+        hint: "An API key from the Perplexity console; the Gateway sends it as a bearer token. Apps POST to /v1/agent for researched answers and /search for result lists, and GET /v1/agent/{id} to poll a background run.",
+        secrets: [
+            {
+                name: "PERPLEXITY_API_KEY",
+                hint: "An API key from the project's API keys page in the Perplexity console.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/orchestrator/v1/models",
+            expect: [200],
+            free: true,
+        },
+        cost: {
+            unit: "token",
+            note: "Agent API per token plus tool fees (web_search $0.0025 a call); Search API $5 per 1,000 requests.",
+        },
+        example: "research the latest pricing changes from our three main competitors",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["search", "ai", "research"],
+    },
+    // r.jina.ai's /echo reflects the request, key included: refused as an account-wide path.
+    jina: {
+        mode: "api_key",
+        label: "Jina AI",
+        description: "Reads web pages as Markdown, searches the web, and embeds and reranks text and images.",
+        auth: {
+            kind: "bearer",
+            secret: "JINA_API_KEY",
+        },
+        docs: "https://jina.ai/api-dashboard/key-manager",
+        api: "https://docs.jina.ai/",
+        hosts: ["api.jina.ai", "r.jina.ai", "s.jina.ai"],
+        data: "per-call",
+        access: ["read"],
+        readMethods: ["GET", "POST"],
+        accountWidePaths: [["v1", "batches"], ["echo"]],
+        hint: 'An API key from the Jina API dashboard; the Gateway sends it as a bearer token. Reader is r.jina.ai (POST {"url": ...}), Search is s.jina.ai, and embeddings and rerank are on api.jina.ai.',
+        secrets: [
+            {
+                name: "JINA_API_KEY",
+                hint: "A key starting jina_, from Manage API Key in the Jina API dashboard.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            host: "api.jina.ai",
+            path: "/v1/batches",
+            expect: [200],
+            free: true,
+            note: "Lists the account's batch jobs, so the Gateway refuses it for apps; run it from the person's terminal.",
+        },
+        cost: {
+            unit: "token",
+            note: "Billed in tokens against the account balance; GET /v1/models lists each model's price.",
+        },
+        rate: {
+            note: "Reader 500 requests a minute, Search 100, embeddings and rerank 500; premium keys get ten times more.",
+        },
+        example: "read this article as Markdown and pull out the key claims",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["scrape", "search", "embeddings", "rerank"],
+    },
+    replicate: {
+        mode: "api_key",
+        label: "Replicate",
+        description: "Runs open-source and official models by API: images, video, audio and language.",
+        auth: {
+            kind: "bearer",
+            secret: "REPLICATE_API_TOKEN",
+        },
+        docs: "https://replicate.com/account/api-tokens",
+        api: "https://replicate.com/docs/reference/http",
+        hosts: ["api.replicate.com"],
+        data: "per-call",
+        access: ["read"],
+        readMethods: ["POST"],
+        accountWidePaths: [
+            ["v1", "account"],
+            ["v1", "deployments"],
+            ["v1", "files"],
+            ["v1", "webhooks"],
+            ["trainings"],
+        ],
+        hint: "An API token from Replicate's API tokens page; the Gateway sends it as a bearer token. Apps POST to /v1/models/{owner}/{name}/predictions or /v1/predictions with Prefer: wait (up to 60 seconds), or pass a webhook URL for longer runs.",
+        secrets: [
+            {
+                name: "REPLICATE_API_TOKEN",
+                hint: "An API token from the API tokens page of the Replicate account.",
+            },
+        ],
+        smoke: {
+            method: "GET",
+            path: "/v1/account",
+            expect: [200],
+            free: true,
+            note: "Account-wide, so the Gateway refuses it for apps; run it from the person's terminal.",
+        },
+        rate: {
+            perMinute: 600,
+            note: "Creating predictions 600 a minute; every other endpoint 3,000 a minute.",
+        },
+        example: "generate a product photo of a ceramic mug on a wooden table",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "images", "video", "audio"],
+    },
+    // The router forwards X-HF-Bill-To as sent, so an app can bill any organization the token's user belongs to (TODO.md).
+    huggingface: {
+        mode: "api_key",
+        label: "Hugging Face Inference",
+        description: "Open models from Hugging Face's inference providers through one router: chat, embeddings, images, speech.",
+        auth: {
+            kind: "bearer",
+            secret: "HUGGINGFACE_API_TOKEN",
+        },
+        docs: "https://huggingface.co/settings/tokens/new?ownUserPermissions=inference.serverless.write&tokenType=fineGrained",
+        api: "https://huggingface.co/docs/inference-providers/index",
+        hosts: ["router.huggingface.co"],
+        data: "per-call",
+        access: ["read"],
+        readMethods: ["GET", "POST"],
+        accountWidePaths: [],
+        hint: "A fine-grained access token from Hugging Face settings with the Inference Providers permission; the Gateway sends it as a bearer token. Apps POST OpenAI-style chat to /v1/chat/completions, or other tasks under /<provider>/, such as /hf-inference/models/<model>.",
+        secrets: [
+            {
+                name: "HUGGINGFACE_API_TOKEN",
+                hint: "A fine-grained token with only the Make calls to Inference Providers permission.",
+            },
+        ],
+        cost: {
+            unit: "call",
+            note: "Pay as you go at each provider's own rate with no markup; PRO accounts get $2 of compute credits a month.",
+        },
+        example: "embed these support articles for semantic search",
+        verified: {
+            on: "2026-10-25",
+            how: "docs",
+        },
+        tags: ["ai", "llm", "embeddings", "images"],
+    },
+};