Multi-tenant ingestion, vector search, and grounded generation behind a simple API. Bring your documents; we handle the retrieval.
Built on serverless infrastructure that scales on demand
The unglamorous parts of a RAG stack — parsing, chunking, embedding, isolation, retries, metering — already built and running in production.
Upload a PDF, Markdown or text file and get an immediate response. A background pipeline then parses, chunks, embeds and upserts it — every phase retried with exponential backoff, every vector id deterministic so re-runs overwrite instead of duplicating.
extract → chunk → embed → upsert → ready
Generation is grounded strictly in retrieved chunks. Emitted [n] markers are resolved back to real sources — with filename, page and score — and anything the model invents is dropped before you ever see it.
Vectors live in a per-tenant, per-collection namespace and every query is filtered by tenant. Another tenant's resources return 404, never a 403 that would leak their existence.
Mint scoped keys from the dashboard, shown once and stored hashed. Traffic is metered per key by an atomic sliding window with RateLimit-* headers on every response.
Queries over time, p95 latency, tokens and cost, plus a failure breakdown across rate limits, empty retrievals and provider errors — no extra instrumentation.
Server-Sent Events by default, so tokens render as they are generated and citations arrive when generation completes. Pass stream: false for one JSON payload that is trivial to evaluate offline.
Every endpoint is generated from the spec, served at /v1/openapi.json, and explorable from a Try it console in the docs. Generate a client in whatever language you ship.
No pipelines to orchestrate, no vector database to operate, no prompt scaffolding to maintain.
Group documents into collections — one per product, customer or knowledge base. Upload PDF, Markdown or plain text up to 25 MB; the call returns immediately.
POST /v1/collections/:id/documentsText is extracted (page numbers intact), split into overlapping chunks, embedded with Gemini and upserted into your namespace. Watch status go processing → ready in the dashboard.
GET /v1/documents/:id/statusYour question is embedded with the same model, retrieved against your namespace, assembled into a bounded context and answered — streamed token by token with validated citations.
POST /v1/collections/:id/queryAuthenticate with an API key, point at a collection, and read back an answer whose every claim maps to a chunk you uploaded.
curl -X POST 'https://rag.lucasfurtado.xyz/api/v1/collections/col_9f8b2a1c/query' \
-H 'Authorization: Bearer rag_live_YOUR_KEY' \
-H 'Content-Type: application/json' \
-d '{"query":"What is the refund policy for annual plans?","stream":false}'{
"answer": "Annual plans can be refunded in full within 30 days of purchase [1].",
"sources": [
{
"marker": 1,
"documentId": "doc_4c1e77a0",
"filename": "handbook.pdf",
"page": 12,
"snippet": "Annual plans are eligible for a full refund within 30 days…",
"score": 0.82,
"cited": true
}
],
"usage": {
"chunksRetrieved": 8,
"chunksUsed": 5,
"contextTokens": 1240,
"invalidMarkers": []
}
}Same API, same isolation guarantees — the collection you point it at is the only thing that changes.
Answer from your help centre and policy docs, with the source article attached to every reply so agents can verify before they send.
Handbooks, runbooks and postmortems become searchable in plain language — scoped per team, so nobody sees a collection they were not granted.
Drop a query box into your docs and return answers with deep links, instead of ten keyword matches the reader still has to skim.
Interrogate long PDFs at page-level precision. Citations carry the page number, so every answer is one click from the paragraph behind it.
Billing is not switched on yet — every tier below is free while the platform is in beta, and existing projects keep their allowances when it lands.
$0forever
Enough to build and ship a first integration.
Customannual
For regulated workloads and dedicated capacity.
Need something in between? Tell us about your workload.
Create a collection, upload a file, and ask it something. The whole loop takes about a minute — no card, no sales call.