Private, grounded document answers

Ask questions of any PDF.

Answers grounded in the document itself, with every claim cited back to the page it came from. Structure-aware chunking, hybrid retrieval, Flashrank reranking, and a quality gate keep the answer tied to the source.

retrieval indexes: dense + sparse

0.65 / 0.35

RRF fusion weighting

±1

neighbour chunks expanded

100%

answers carry citations

The problem

A plain LLM will confidently invent what your document says.

It has never read your file. Ask it about a 200-page report and it produces something plausible, unattributable, and occasionally wrong in the exact places that matter.

The approach

Retrieve first. Answer only from what was retrieved. Cite it.

The document is split around its own structure, scored for quality, and indexed twice: dense vectors and sparse keywords. Every question is rewritten against recent chat, fused, reranked, and handed only the surviving chunks.

How it works

Illustrative view of the backend pipeline, not measured telemetry.Hover a stage for a plain-language explanation.

Security by default

The boundaries around the pipeline matter as much as the models.

Account-scoped access

JWT authentication is required for ingest, query, documents, and chat history.

Document isolation

Every vector search is filtered by the signed-in owner and selected document.

Untrusted-data boundary

Document text is context, never instructions; guardrails keep retrieved content from steering the system.

Abuse controls

Per-user rate limits protect the expensive ingest and query paths.

Key handling

The optional BYO OpenAI key is sent in a request header and is never persisted by the backend.

Architecture

API

FastAPI

Ingest, query, health and collection endpoints. Every response carries an X-Request-ID.

Vectors

Qdrant

Qdrant Cloud stores document-scoped vectors with owner and document filters.

Keywords

BM25

Sparse index running alongside the vectors, catching exact terms embeddings miss.

Reranker

Flashrank

Re-scores the fused candidate set before anything reaches the model.

Models

Azure + BYO

Azure-hosted gpt-5-mini by default; an optional OpenAI key is used only for your request.

Runtime

Azure Container Apps

Scale-to-zero hosted runtime; Docker Compose remains available for local development.

What's in the pipeline

Structure-aware chunking

Splits around headings and sections, falling back to recursive chunks only when a section is oversized.

Chunk quality gate

Scores chunks and applies a soft retrieval penalty; every non-empty chunk remains indexed.

Hybrid retrieval

Dense vectors and sparse keywords searched in parallel.

RRF fusion

Reciprocal Rank Fusion merges both rankings (0.65 dense, 0.35 sparse).

Reranking

Flashrank re-orders candidates by true relevance to the question.

Neighbour expansion

Pulls adjacent chunks so answers keep their context.

Inline citations

Every claim tagged with its file and page, rendered as linked chips.

Parser router

Routes PDF, Office, image and CSV files through local and OCR adapters with fallback tiers.

JWT + guardrails

Every document route is account-scoped and retrieved text is treated as untrusted data.

RAGAS harness

Faithfulness and relevancy measured offline, not guessed at.

Where your data goes

Your account keeps its own documents and chat history; the default model runs through our Azure deployment.

Retrieval uses Qdrant Cloud and self-hosted bge-m3 embeddings. Generation uses Azure-hosted gpt-5-mini by default. If you provide an OpenAI key, it is sent per request and never stored on the backend.

Frequently asked

Is my data private?

Yes. Every document, embedding, and chat history is scoped to your account — retrieval is filtered by owner on every query, so another user can never see or search your documents.

Do you use my documents to train any model?

No. Documents are used only to answer your own questions about them. Nothing you upload is used for training.

What happens when I delete a document?

Deleting a document removes its database record, its vectors from storage, and its retrieval index — it isn't just hidden from the list.

Which file types are supported?

PDF, TXT, Markdown, CSV, JSON, Word, PowerPoint, Excel, and common image formats (PNG, JPG, WEBP), routed through local parsing or an OCR fallback depending on the file.

Can I use my own OpenAI key?

Yes, from the workbench. It swaps the model that generates answers; retrieval always uses the same embedding model a document was indexed with. Your key is sent with the request and never stored on the backend.