Data Greffe Justice
A full stack RAG system processing 50,000+ French legal documents: triple index Elasticsearch search, three tier RBAC, Claude powered anonymization, OHADA compliance, and ~23 documented endpoints.
- Documents
- 50K+
- PII accuracy
- 98%
- Response
- 2s
Selected engineering work by Muhammad Mubashir, showcased by Retrieva Labs. Figures and scope reflect the existing project notes. The evidence section explains what is available to review.
A useful outcome starts here.
A legal document platform needed semantic retrieval, structured search and answers that retain their source context.
From input to useful output.
- 01Extract and structure documents
- 02Anonymize sensitive fields
- 03Build search indexes
- 04Retrieve relevant passages
- 05Generate an answer with sources
The choices behind the system.
Several search needs
Separate Elasticsearch indexes support semantic search, metadata filtering and decision chains. Pinecone provides the vector search layer described in the notes.
Context before generation
Chunking and hybrid retrieval put relevant document passages in context before Claude generates an answer.
Evaluate the pipeline
The engineering note describes comparing chunk sizes against a set of 200 legal question and answer pairs prepared with a lawyer.
Project notes, with context.
Project notes report a corpus of 50,000+ documents and anonymization and latency results. The underlying evaluation data is not public. Retrieval confidence scores are not a measure of retrieval accuracy.
Read the engineering note