Direction 2 · RAG Knowledge Bases¶
The value of RAG isn't the vector DB — it's trustworthy citation + versioned lineage + measurable retrieval. 3 representative cases + delivery discipline.
Core Capabilities¶
- Embedding: bge-large · bge-m3 · text-embedding-3 · fine-tuning
- Reranking: BGE-reranker · Cohere Rerank · Cross-Encoder · MMR
- Vector DB: Milvus · Qdrant · Chroma · pgvector · Elasticsearch KNN
- Document processing: OCR · layout parsing · table extraction · chapter chunking · version-aware
- Evaluation: Ragas · TruLens · manual labels · hallucination rate · citation accuracy
3 Representative Cases¶
Case 01 · Law Firm Tri-Corpus RAG (Case Law + Contracts + Regulations)¶
- Problem: lawyer clause search averaged 45 minutes · versions confused · citation errors delivered
- Stack: bge-large embedding · BGE-reranker · tri-corpus tiered retrieval · clause-level citation traceability · version-aware routing
- Boundary: client contract isolated tenants · read-only chain
- KPI: lawyer clause search P50 45min → 3min · citation accuracy >95% · lawyer active-use rate >70%
- My role: RAG architecture · tri-corpus tiering strategy · embedding fine-tuning · lawyer joint design
Case 02 · Power Grid Field Voice-Query RAG¶
- Problem: high-noise substation field maintenance · manuals cumbersome · incident experience fragmented
- Stack: whisper-large voice entry + tri-source RAG (manuals + incidents + SOPs) + mobile app
- Boundary: field equipment offline queryable · incident archives internally authorized
- KPI: maintenance-ticket filing -55% · field query time 10 min → 30s · new-hire learning curve -40%
- My role: voice pipeline · tri-source fusion · on-site deployment
Case 03 · Medical Device Support KB (Tier-1 Ticket -40%)¶
- Problem: customer-support tier-1 slow · tier-3 experts flooded
- Stack: bge-large + BGE-reranker · versioned product docs · ticket-feedback loop retraining
- Boundary: post-market feedback separate corpus
- KPI: tier-1 ticket response -40% · tier-3 expert escalation -68% · CSAT +12%
- My role: RAG architecture · embedding fine-tuning · doc version strategy
Delivery Discipline¶
- Doc processing is the heavy lift: OCR · table extraction · chapter chunking · version-awareness — 60% of engineering
- Don't jump to premium embeddings first: start with bge-large; fine-tune only if needed
- Reranking is mandatory: pure embedding top-k has 50-80% hallucination · adding rerank drops to 5-10%
- Citation traceability: each chunk records source doc + section + version + timestamp · output must include citation
- Continuous evaluation: build 100-500-item eval set · regression on every index refresh
Common Anti-Patterns¶
- ❌ Dump PDF straight into embedding without OCR / table extraction — index quality abysmal
- ❌ GPT for QA only, no reranking — severe hallucination
- ❌ No version awareness — old regulations still surface after updates
- ❌ Milvus deployed but query 500ms+ — no ANN parameter tuning
Related¶
- Full 20-case portfolio · PDF
- Related directions: Private LLM deployment · Agent workflows
Next: Download PDF · 30-min RAG alignment