Auto-Synchronized RAG

Problem

For lawyers, judges, and researchers, legal answers lived inside voluminous PDFs spread across diverse sources — and a generic chatbot would only make things worse: fluent, confident, and wrong about clauses. The brief demanded answers grounded in the institution's own documents: always sourced, always current.

• Manual consultation too slow across large, complex corpora
• Relevant passages hard to locate precisely
• No source traceability on generated answers
• No quality tracking or consultation history

Context

Two constraints shaped everything. Juridical documents are deeply hierarchical — headings, articles, and clauses carry the meaning — across languages and layouts, so flattening them destroys retrievability. And the corpus is living: texts evolve constantly. The system had to guarantee reliability, traceability, and currency at once, while keeping update costs proportional to change rather than corpus size.

Approach

Three pipelines with clear responsibilities, custom-engineered without heavy orchestration frameworks for full control of sync state, memory, and latency.

• Ingestion + synchronization — detects added, modified, and removed documents; indexes only what changed, never a full re-index
• Retrieval — semantic search combined with a reranking stage to isolate the most relevant passages
• Generation — answers anchored exclusively in retrieved context, in the question's language, with sources attached or an explicit refusal

Architecture

A vector-backed knowledge store behind the auto-synchronization layer, served by an API with real-time streaming and multi-user sessions — containerized for reproducible deployment.

• Knowledge store with optimized similarity index, converged automatically with the document folder
• Standard + streaming chat endpoints, ratings, session control, and secure document serving
• Three roles by design: end users, ingestion administrators, downstream systems via REST
• Every answer written to an audit store; analytics service reports volume, approval rate, and per-response detail

Implementation

What makes it production rather than a demo is the quality loop: usage feeds directly back into measurable answer quality.

• Incremental ingestion — unchanged knowledge is never re-processed
• Streaming interface with typing state, expandable source links, and per-answer feedback
• Follow-ups and ratings surfaced only on grounded answers
• Admin analytics: KPIs, rating / date / keyword filters, volume trends, per-session rollups, full query-answer audit

Challenges

Each failure mode of naive RAG got a dedicated fix — not a prompt tweak.

• Complex layouts broke extraction — high-accuracy pass preserving order and hierarchy
• Fragmented passages hurt recall — structure-aware chunking with heading context
• Plausible near-misses in search — dedicated reranking stage with graceful degradation
• History-vs-document confusion caused false citations — grounding rule separating chat history from document context

Outcome

Deployed and validated against its objectives — functional, performance, and evaluation layers all green. In daily use: drop a file in the folder and the index converges on its own; ask in plain language; get a fast, sourced answer — and audit everything afterward.

• Functional tests pass: 10+ juridical PDFs, multilingual queries, real-time streaming, multi-user sessions, ratings, analytics
• Performance within target: answers typically generated in 2–4 seconds with concurrent use supported
• Evaluation harness: 100% retrieval hit rate and recall, 96.7% precision at K, 88% faithfulness, 90% contextual recall, 80% correct refusal