Retrieval-Augmented Document Intelligence

Talk to your documents.
Get answers you can trust.

DocuChunk is a full RAG platform for interacting with your documents precisely and in context. Upload a file, ask in plain language, and get a grounded, cited answer — backed by hybrid semantic retrieval, cross-encoder reranking, and source-level verification. Your data stays private to your workspace.

Start free See how it works
Hybrid
Dense + BM25 retrieval
Cited
Every claim traced to source
Private
Per-user isolation
“Which elf is the accountant, and who supervises the fudge shoppe?”
Flo works as the accountant [1], while Zack supervises the fudge shoppe [2].
✓ Grounded & verified
PDF · DOCX
Supported formats
384-dim
Semantic embeddings
<10ms
Vector retrieval
Cited
Grounded answers
Capabilities

A complete retrieval & reasoning stack

Every layer from raw file to trustworthy answer — engineered, not glued together.

Hybrid semantic search
Dense vector similarity fused with BM25 keyword matching via Reciprocal Rank Fusion — so exact entity names and fuzzy meaning both surface the right passage.
Cross-encoder reranking
A joint (query, passage) reranker re-scores candidates to reject topically-similar but wrong-entity matches — the precision layer bi-encoders miss.
Grounded & cited answers
Every claim carries a [n] citation to its source passage, and a separate verification pass checks the answer against the sources — no silent hallucinations.
Scope-aware answers
Ask for a fact, a full list, a table, a classification, or a summary — the router adapts retrieval to the question, reading the whole document when the answer spans it.
Embedding visualiser
A UMAP-projected 2D map shows how your chunks cluster in embedding space — inspect any chunk's vector and its nearest neighbours to understand retrieval.
Streaming, in-context chat
Answers stream token-by-token, then a trust signal (confidence + citations) lands a moment later. Ask follow-ups in a conversation tied to the document.
Pipeline

From upload to grounded answer

Ingestion runs automatically in the background; retrieval and generation run on every question.

01
Parse & clean
Text is extracted page-by-page, then normalised — ligatures fixed, hyphenated breaks joined, header/footer noise stripped.
background worker
02
Chunk with overlap
Sentence-aware chunking with overlap and stored character offsets, so answers that straddle a boundary can be stitched back together.
sentence-aware
03
Embed & index
Each chunk becomes a 384-dim vector (all-MiniLM-L6-v2), stored in an HNSW index and mirrored in a per-user keyword index.
ChromaDB · BM25
04
Retrieve & rerank
Hybrid retrieval fuses dense + keyword hits with RRF, then a cross-encoder reranks to the passages that truly answer the query.
RRF · cross-encoder
05
Generate with citations
The reranked passages are assembled into a grounded prompt; the model answers in its own words with inline [n] citations.
grounded prompt
06
Verify & score
A separate pass checks each claim against its cited source and returns a confidence breakdown — retrieval, citation coverage, groundedness.
verification
Private by design

Your documents stay yours

DocuChunk is built around isolation and traceability — not a black box you have to trust blindly.

Per-user isolation
Each account gets its own vector collection and keyword index — retrieval never crosses users.
Self-hostable stack
Runs entirely on open components you control — no vendor lock-in, deployable on your own box.
Traceable answers
Every answer links back to the exact source passage and page, so you can verify it yourself.
Secure auth
JWT sessions, bcrypt-hashed passwords, TLS in transit, optional OAuth sign-in.
Workspace: yours only
collection: user_••• · isolated
quarterly_report.pdfindexed
contract_v3.docxindexed
research_notes.pdfprocessing…
Open stack

Built on best-in-class open tools

Every component is open and swappable — deploy it your way.

FastAPI
API framework
ChromaDB
Vector store · HNSW
Sentence-Transformers
Embeddings · MiniLM
Cross-Encoder
Reranking
Gemini
Grounded generation
UMAP
Embedding visualiser
PyMuPDF
Document parsing
JWT · OAuth
Authentication
Get started

Put your documents to work

Create a free account, upload a file, and ask your first question in under a minute — with grounded, cited answers from the start.

Create free account Sign in