Ask once. Get the evidence.
Create a key, upload a file, and search. Each key accesses only its own workspace.
1. Create a key
Sign in, add credit from $10, and create an API key in Profile. Keep it on your server as EMBEDDINGS_API_KEY.
2. Upload a file
curl https://api.embeddings.app/v1/documents \
-H "Authorization: Bearer $EMBEDDINGS_API_KEY" \
-F "file=@Company-policy.pdf" \
-F "model=text-embedding-3-small"PDF, DOCX and UTF-8 text (TXT, Markdown, CSV, JSON, HTML and XML) are processed into searchable passages. PDFs must contain selectable text; native visual models can accept supported PDFs or page images after activation. File limit: 5 MB, 100 PDF pages and 60 KB of extracted text. Each workspace model index supports 5,000 passages.
GET /v1/documents?limit=20&cursor=RETURNED_CURSOR
GET /v1/documents/{id}
GET /v1/documents/{id}/content
PUT /v1/documents/{id} (replacement multipart file)
DELETE /v1/documents/{id} (recoverable removal)
POST /v1/documents/{id}/restore {}Use the returned next_cursor for the next page; limit is 1–50. Upload responses include processing status, model and dimensions. JSON {name,text,model} is also supported. Removed and failed files are excluded from search. Replacement excludes old passages immediately. Files are processed in the background; check their status before searching.
3. Search
curl https://api.embeddings.app/v1/search \
-H "Authorization: Bearer $EMBEDDINGS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"query":"What are the contract termination terms?"}'Returns ranked source passages with file name, source URL and page or line where available. Results combine semantic and keyword relevance. This service returns evidence; it does not generate an answer.
{
"model": "YOUR_SELECTED_MODEL",
"results": [{
"rank": 1,
"document_id": "...",
"text": "The relevant source passage...",
"source": {"name": "Contract.pdf", "page": 3, "url": "/v1/documents/.../content"}
}]
}Optional fields: model overrides your workspace selection and top_k selects 1–10 passages (default 5). Relative source URLs use the API host and require the same bearer key. Scores order results and are not calibrated probabilities.
Choose your model
Every current model appears in the selector, with its activation state. Choose in Dashboard or use the API:
GET /v1/search/models
POST /v1/search/settings
{"model":"text-embedding-3-large"}New uploads and query-only searches use your selection. Each file keeps its indexed model. A model change never starts automatic re-indexing. Use “Use selected model” on a file to explicitly re-index it; new inputs use prepaid credit.
Different model spaces are kept separate. Search only returns files indexed with the selected model. Models awaiting activation do not dispatch provider calls. Historical text model IDs remain compatible with the embedding API.
MCP
Connect a Streamable HTTP MCP client to https://api.embeddings.app/mcp with Authorization: Bearer YOUR_API_KEY. The server exposes one tool:
search({"query":"Find the early termination clause"})Stateless JSON-RPC over HTTP POST. Supports initialization, tool discovery and calls with protocol version 2025-06-18. Results include structured source evidence. Optional model and top_k match the Search API. No browser session or generated answer is required.
Compatible vector API
Existing integrations can continue to use the OpenAI-compatible embedding endpoint:
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.EMBEDDINGS_API_KEY,
baseURL: "https://api.embeddings.app/v1",
});
const vectors = await client.embeddings.create({
model: "text-embedding-3-small", input: ["Workspace pulse", "收件箱"],
});GET /v1/models lists active text models; POST /v1/multimodal/embeddings accepts bounded native media. Direct vectors are optional: file search handles parsing, chunking, indexing and reranking for you.
Images, audio and video
Choose an activated native model in Dashboard, then upload PNG/JPEG, WAV/MP3, MP4/MOV or a compatible PDF. See each model guide for supported types. Media access awaits commercial activation and has a separate request price.
Include duration_seconds for audio/video uploads. Each native media request produces one combined vector. Encoded input is limited to 1 MiB or the model’s stricter limit. Default clips are up to 30 seconds; Gemini supports 180-second audio and 120-second video. Large recordings must be split before upload. Search uses a text query in the same model space.
Usage and limits
Included during early access: 100 MB of originals, 1,000 files, 5,000 passages per model and 1,000 search requests per workspace per UTC month. Shared service budgets also apply. Contact support before larger imports.
Removed files can be restored for 7 days. Eligible originals and index entries are cleared during maintenance. Cached vectors and billing records have separate retention. Scanned PDF OCR, Excel formulas and large media processing are not included.
New text inputs use prepaid credit. EmbeddingGemma 2 costs $0.02 per million actual input tokens; other text models use their published UTF-8 input-byte price. Identical normalized inputs reuse cached vectors for free. Media uses its separate activated request price. GET /v1/usage returns the balance and recent operations.
401: check your key. 402: add credit. 429: rate or daily budget limit. Pending activation returns 503 before dispatch. Unknown provider outcomes are held for reconciliation; do not retry or upload under a different model to work around a held call. Contact support to reconcile. Original files remain private and source downloads always require authorization.