RAG Architect

Design systems that connect LLMs to your private data (Retrieve -> Augment -> Generate).

When to Use This Skill

• Building "Chat with your Data" applications
• Selecting a Vector Database (Pinecone, Milvus, pgvector)
• Designing Chunking Strategies (Fixed-size, Semantic, Markdown)
• Hybrid Search (Keyword + Semantic) implementation
• Reducing hallucinations with grounded context

Core Concepts

1. Ingestion Pipeline (The "Loader")

Garbage In, Chatbot Out.

1. Load: Extract text from PDF, HTML, Notion.
2. Chunk: Break into pieces.
   • Fixed: 500 chars (Simple, breaks context).
   • Semantic: Break by paragraphs/headers (Smart).
3. Embed: Convert text to vectors (OpenAI text-embedding-3-small, Cohere).

2. Retrieval Strategies

Vector Search (Semantic)

• Finds "meaning". ("Can I bring my dog?" matches "Pet Policy").
• Pros: Understands intent.
• Cons: Fails on exact keywords (IDs, acronyms).

Keyword Search (Lexical)

• BM25 / TF-IDF.
• Pros: Precise for exact terms.
• Cons: Dumb.

Hybrid Search (The Gold Standard)

• Combine Vector + Keyword.
• Use Reciprocal Rank Fusion (RRF) to re-rank results.

3. Generation (The "Synthesis")

Construct the final prompt.

Answer the question based ONLY on the following context:
{retrieved_chunks}

Question: {user_query}

Advanced RAG

• HyDE (Hypothetical Document Embeddings): Generate a fake answer, embed that, then search.
• Parent Document Retriever: Search small chunks, but return the big parent chunk to the LLM.
• Re-ranking: Use a Cross-Encoder (Cohere Rerank) to score the top 20 retrieved docs before sending to LLM.

Resources

• LangChain RAG Docs
• Building RAG from Scratch