Retrieval & grounded inference

Legal document RAG

A production retrieval-augmented generation pipeline over more than 100,000 legal files.

  • Backend
  • AI inference

The context

A large legal document collection needed an application pipeline that could retrieve relevant sources and construct grounded LLM responses. The work covered the path from document ingestion to generation.

My contribution

  • Implemented ingestion, chunking, embeddings, and retrieval over 100,000+ legal files.
  • Built context-construction and generation workflows and integrated them into the client application.
  • Designed retrieval and grounding strategies around relevance, context utilization, latency, and unsupported-response reduction.

Challenges & solutions

Making a large legal collection usable as context.

Implemented document ingestion, chunking, embedding, and retrieval stages rather than passing the full collection to a model.

Constructing relevant, grounded model context.

Designed source selection and context-construction strategies with attention to relevance and the available context window.

Pipeline overview

A simplified view of the workflow.

  1. Legal files
  2. Chunking & embeddings
  3. Relevant retrieval
  4. Context & generation

Scope & outcomes

Legal files in the system
100K+
Ingestion through generation
End-to-end

Client documents, identifying details, and source code are not shared. No quantitative retrieval-accuracy or hallucination-reduction benchmark is claimed.

Building something related?

Tell me about the problem, the constraints, and what success would look like.

Discuss a similar project

Explore more work

Local AI runtime & agent orchestration

Localbots

A Go-based local agent framework and web studio: a manager coordinates specialist bots through one resident Qwen3 model, with no cloud API required after setup.

Backend & synchronization

Smartlistzero

Go-powered CRM synchronization and performance analytics across 70+ agencies managing approximately 5 million contacts.