Enterprise RAG (Retrieval-Augmented Generation) Architecture

System Directive • Copy and Paste Below

Act as a Principal AI Engineer specializing in production Retrieval-Augmented Generation (RAG) architectures and vector retrieval.

Design an end-to-end, high-accuracy RAG pipeline for our enterprise internal documentation and knowledge base.

### CORPUS & REQUIREMENTS:
- Source Data Type (e.g. PDF manuals, Notion docs, SQL database, Confluence, Zendesk tickets): [SOURCE_DATA_FORMATS]
- Approximate Volume & Update Frequency: [VOLUME_AND_UPDATE_FREQUENCY]
- Target User Queries & SLA Latency: [USER_QUERY_TYPES_AND_LATENCY]
- Chosen LLM & Embedding Model (e.g. text-embedding-3-large, Cohere, Claude 3.5, GPT-4o): [MODELS_SELECTED]
- Target Vector Database (e.g. Pinecone, Supabase pgvector, Qdrant, Milvus): [VECTOR_DATABASE]

### SYSTEM BLUEPRINT:
1. **Ingestion & Preprocessing Pipeline**:
   - Document cleaning, metadata extraction (author, document type, security tier, last updated).
   - Chunking Strategy: Compare Fixed-size vs Semantic vs Recursive chunking with overlap parameters for this specific data type.
2. **Hybrid Search Architecture**:
   - Dense Vector Retrieval (semantic similarity) + Sparse Keyword Retrieval (BM25 for exact codes, SKUs, and proper nouns).
   - Reciprocal Rank Fusion (RRF) formula to combine scores.
3. **Re-Ranking & Context Window Optimization**:
   - Re-ranker model selection (e.g. Cohere Re-ranker) to compress top 20 candidate chunks down to top 4 most relevant chunks.
   - Context stuffing prevention and token count management.
4. **Prompt Engineering & Hallucination Guardrails**:
   - Production system prompt enforcing strict citation, ground truth boundaries, and graceful refusal when answers are not in context.
5. **Evaluation & Drift Monitoring**:
   - Ragas / TruLens evaluation metrics: Faithfulness, Answer Relevance, and Context Precision.

More Prompts in AI Automation & n8n