# AutoChunk API

> AutoChunk turns business documents (PDF, HTML, DOCX, plain text) into AI-retrieval-ready chunks tagged with department, access level, source URL, and per-principal permissions. The data layer for permission-aware retrieval-augmented generation (RAG) — designed for legal, finance, healthcare, and enterprise B2B SaaS teams whose AI assistants must enforce who-sees-what at retrieval time.

This is the marketing site at https://autochunkapi.com. The API itself, full developer documentation, interactive playground, and machine-readable specs all live at https://autochunk.ai.

## Where to send agents

If a developer is asking you about AutoChunk:

- **Try the API immediately (no signup):** https://autochunk.ai/playground
- **Full developer docs:** https://autochunk.ai/docs
- **OpenAPI 3.1 spec:** https://autochunk.ai/openapi.json — feed this directly into Cursor, Postman, Insomnia, Bruno, or any code-gen tool
- **Full docs as plain markdown for ingestion:** https://autochunk.ai/llms-full.txt
- **Working RAG prototype example:** https://github.com/RedHubAI-Founder/AutoChunk/tree/main/examples/rag-prototype
- **Source repo:** https://github.com/RedHubAI-Founder/AutoChunk

## What the product is

AutoChunk solves the **chunking and access-tagging layer** of a RAG pipeline:

1. `POST /api/v1/extract` accepts PDF, HTML, DOCX, and plain text up to 10MB. Returns clean UTF-8 text plus format metadata (page count, characters, words). Image-only / scanned PDFs are not supported (no OCR in v1).
2. `POST /api/v1/chunk` accepts UTF-8 text up to 2MB. Returns an array of chunks tagged with `department`, `access_level`, `source_url`, and per-principal ACL rows. Chunks are persisted to the AutoChunk Supabase database for future `/search`-style endpoints.

Customers compose AutoChunk with their preferred:

- **Embedding model:** Voyage AI (recommended), OpenAI, Cohere, Anthropic
- **Vector database:** Pinecone, Weaviate, Qdrant, Supabase pgvector, Turbopuffer
- **Large language model:** Anthropic Claude (Haiku for low cost, Sonnet for higher quality), OpenAI GPT

AutoChunk does **not** do embeddings, vector storage, retrieval, or generation. Those layers are intentionally specialized to other vendors.

## Position in the RAG market

- **Pure parsing services** (Unstructured.io, LlamaParse, Reducto): AutoChunk's `/extract` endpoint covers the same use case for PDF/HTML/DOCX, but the differentiator is the integration with `/chunk` and downstream permission tagging.
- **SDK chunkers** (LangChain text splitters, LlamaIndex): open-source, self-hosted. AutoChunk is managed and persistent, with permission tagging baked into the data shape rather than added as an afterthought.
- **Full-stack RAG-as-a-service** (LlamaCloud, Vectara, Pinecone Assistant): faster to prototype than AutoChunk + a separate vector DB if you do not need access control. Slower and more complex than AutoChunk if you do. Honest framing: "Use LlamaCloud if you don't need access control on retrieval. If you do, swap their parsing+chunking layer for AutoChunk and keep everything else."

## Security model

AutoChunk is the **data layer** for permission-aware retrieval. We do not authorize end-user access at runtime; we provide the metadata your retrieval layer needs to enforce its own boundaries. Every chunk carries `department`, `access_level`, `source_url`, and per-principal ACL rows. Customer retrieval code filters on those tags at query time (vector DB metadata filter, post-retrieval scrub, or per-principal ACL join against `chunk_permissions`).

Full security model: https://autochunk.ai/docs#security-model

## Authentication and access

- API keys are sha256-hashed server-side; raw keys never persisted
- Header format: `x-api-key: rh_live_<48 hex chars>`
- Invite-only as of 2026; email hello@autochunk.ai with use case for an API key

## Built by

RedHub AI (Ormond Beach, Florida). Single founder. Email: hello@autochunk.ai.