RAG Pipeline Design

The Basic RAG Pipeline

1. Ingestion: Load documents → Chunk → Embed → Store in vector DB
2. Retrieval: Embed user query → Search vector DB → Get top-K relevant chunks
3. Generation: Construct prompt with retrieved context + user question → Send to LLM → Return answer

Retrieval Prompt Template

Answer the user's question based on the following context.
If the context doesn't contain enough information, say so.

Context:
{retrieved_chunks}

Question: {user_question}

Answer:

Advanced RAG Techniques

Query Transformation

• Query rewriting: Use an LLM to reformulate the query for better retrieval
• HyDE: Generate a hypothetical answer, embed that instead of the question
• Multi-query: Generate multiple query variants and combine results

Re-ranking

After initial retrieval, use a cross-encoder model to re-rank results for relevance. More accurate than embedding similarity alone.

Agentic RAG

Use an AI agent that can iteratively search, evaluate results, and refine queries until it finds sufficient information.

Evaluation Metrics

• Retrieval: Precision@K, Recall@K, MRR (Mean Reciprocal Rank)
• Generation: Faithfulness (does the answer match the context?), Relevance, Completeness
• Tools: RAGAS, DeepEval, LangSmith for automated RAG evaluation

Common Pitfalls

• Chunks too large → noisy, irrelevant context fills the prompt
• Chunks too small → missing context, fragmented information
• No metadata filtering → retrieving outdated or irrelevant documents
• Ignoring evaluation → no way to know if changes improve quality

🌼 Daisy+ in Action: Grounded AI Responses

Daisy+ implements a practical RAG pipeline: incoming customer questions on livechat trigger a search against product docs and FAQ content, the relevant chunks are injected into the LLM's context, and the response is grounded in actual company data rather than hallucinated answers. This is why DaisyBot can answer "what's your return policy?" accurately — it retrieves the real policy, not a guess.