01 / PROBLEM STATEMENT
Why this work exists
Financial and regulatory documents can be scanned, multilingual, and inappropriate for wholesale external indexing. The prototype keeps extraction and retrieval local while sending only retrieved context to generation.
Data and source material
- PDF documents with native extraction and Tesseract OCR fallback
- Semantic chunks with overlap
- Multilingual MiniLM embeddings stored in a FAISS index
02 / SOLUTION OVERVIEW
From input to usable output
Extract locally, chunk semantically, index embeddings in FAISS and send only retrieved context to generation.
- 01
Extract PDF text or invoke French OCR
- 02
Split content into overlapping semantic chunks
- 03
Embed and index chunks locally
- 04
Retrieve relevant context for a question
- 05
Inject bounded context into deterministic DeepSeek generation through FastAPI
03 / TECHNICAL ARCHITECTURE
Implemented components
PDF/OCR → chunks → embeddings + FAISS → context → LLM via FastAPI
04 / METHODOLOGY & IMPLEMENTATION
Decisions that shape the system
- paraphrase-multilingual-MiniLM-L12-v2 creates multilingual embeddings.
- FAISS performs local vector similarity search.
- The generation configuration uses zero temperature, while SlowAPI rate limiting protects the web endpoint.
Implemented capabilities
- PDF + OCR
- Multilingual embeddings
- FAISS
- Context grounding
- Rate limiting
05 / RESULTS & EVIDENCE
What the source supports
- The repository documents a working local retrieval and FastAPI prototype with a browser interface.
- It does not provide a retrieval benchmark, answer-accuracy evaluation, or production-security assessment.
Results preserve the repository’s units, evaluation framing, and stated limitations. Missing evidence is not converted into a claim.
06 / CHALLENGES & LIMITATIONS
Where the evidence stops
- An external DeepSeek key is required for generation.
- Zero temperature improves repeatability but cannot guarantee factual answers.
- The public documentation does not quantify retrieval recall, hallucination rate, latency, or OCR quality.
07 / WHAT THIS DEMONSTRATES
Engineering signal
Document ingestion, OCR fallback, semantic retrieval, API design, and privacy-aware separation of local processing from external generation.
08 / SOURCE & RELATED WORK
Trace the evidence
d44742c80a193d75d444726b9dc30d1557fbe543 ↗README.md · main.py · pdf_reader.py · recherche_semantique.py · prompt.py