Document ingestion
- 01Validate upload
- 02Extract text
- 03Chunk deterministically
- 04Create embeddings
- 05Index in Qdrant
RAG DOCUMENT INTELLIGENCE / LOCAL-FIRST MVP
ATLASTASK THE ARCHIVE. VERIFY THE SOURCE.
Turn a private document archive into answers you can inspect.
ATLAST ingests local PDF and TXT files, retrieves the most relevant passages, asks a local Ollama model to answer from that evidence, and exposes only citations validated against the retrieved sources.
01 / PROBLEM & VALUE
Useful answers need visible evidence.
THE PROBLEM
Operational knowledge is trapped across PDFs and text files. Keyword search finds strings, but rarely gives a concise answer with enough context to trust it.
THE SOLUTION
Build a deterministic local index, retrieve passages by semantic similarity, bound the evidence sent to a local model, then validate every returned citation ID.
THE VALUE
A usable document-intelligence workflow whose source passages, document titles and page or section metadata stay inspectable from answer to archive.
02 / LIVE SYSTEM FLOW
Two observable paths through one local system.
The interface confirms request-level stages from real API outcomes and returned metadata. It does not claim streamed, real-time backend events or invented timings.
Grounded question answering
- 01Retrieve passages
- 02Select evidence
- 03Assemble bounded context
- 04Generate with Ollama
- 05Validate citations
- 06Return answer
03 / WORKING FEATURES
A complete local document lifecycle.
PDF & TXT ingestion
Extract text and retain stable document identity, filename, page or section metadata where available.
Deterministic indexing
Content hashes detect duplicates; deterministic chunks and lightweight Sentence Transformers embeddings feed persistent Qdrant local storage.
Grounded answers
Document text is treated as untrusted data. The prompt constrains generation to retrieved evidence and unsupported questions can return an insufficient-evidence response.
Validated citations
Citation IDs are accepted only when they match retrieved passages, with title and page or section metadata returned alongside them.
Lifecycle management
List, replace and delete one document at a time. Replacement keeps the existing document until the new upload indexes successfully.
CLI, API & web UI
The same ingestion, retrieval and QA services power command-line workflows, FastAPI endpoints and the responsive browser interface.
04 / ARCHITECTURE
Small components. Explicit trust boundaries.
A buildless web interface calls a loopback FastAPI application. The backend owns validation, extraction, embeddings, Qdrant persistence, bounded retrieval context, Ollama generation and citation validation.
BROWSER → FASTAPI → QDRANT ↔ OLLAMA
- Interface
- Responsive HTML, CSS and JavaScript for uploads, the document archive, system-flow status, answers and inspectable citations.
- API
- FastAPI endpoints for health, upload, listing, replacement, deletion, retrieval and grounded question answering.
- Ingestion
- Python extraction for PDF and TXT, SHA-256 document identity and deterministic word-window chunking.
- Retrieval
- Sentence Transformers embeddings with Qdrant local mode for persistent vector search.
- Generation
- A configurable local Ollama model receives bounded evidence and an injection-resistant grounding prompt.
- Validation
- Returned citation IDs are checked against the retrieved passage set before the structured response is exposed.
05 / RELEASE VERIFICATION
Verified release evidence, with a narrow claim.
A6 release verification used the packaged wheel and a real installed Ollama gpt-oss:20b model. The retrieval benchmark contains six labeled, answerable questions in a small synthetic corpus.
- Automated tests
- 40 passing
- Recall@3
- 1.0
- MRR
- 1.0
- Answerable questions
- 6 synthetic
These perfect synthetic retrieval scores do not establish general RAG accuracy, answer quality or performance on real organizational documents.
06 / REAL SCREENS
The working interface, from archive to evidence.
All five screens come from the local application using synthetic documents. There is no hosted interactive demo.
07 / LOCAL EXECUTION
Run the real application locally.
ATLAST is local software, not a public API. It requires Python 3.12, uv and a locally installed Ollama model. The default embedding model is downloaded on first use and cached locally; ATLAST never downloads an Ollama model automatically.
git clone https://github.com/fadel-santiago/atlast.git
cd atlast
uv python install 3.12
uv sync --frozen
uv run --frozen atlast --db .atlast/qdrant ingest examples/sample.txt
uv run --frozen atlast --db .atlast/qdrant serve --ollama-model gpt-oss:20b --port 800008 / LIMITATIONS