Retrieving RAG Context

Retrieving RAG Context

Process for querying the Qdrant vector store to provide rich context for AI responses using an Agentic RAG approach (Retrieve -> Grade -> Adapt).

Prerequisites

• Active Qdrant vector store with indexed documents (running via Docker).
• FastMCP RAG Server must be actively running via scripts/mcp/servers/rag/start_rag_server.bat (exposing SSE on port 8000).
• Knowledge of the target ebook library structure.

When to Use

• When answering complex technical or domain-specific questions.
• When grounding responses in project-specific documentation or ingested ebooks.

Process

0. Initialize State:
   • Before beginning, create an internal memory buffer (rag_search_state) to accumulate all relevant retrieved information.
1. Retrieve (Formulate Query & Search):
   • Formulate a precise, domain-specific search query based on the user's need.
   • TOC Reconnaissance (Optional but Recommended): Every ingested eBook has a deterministic outline injected into the vector store. If you need to understand the broad structure of the book before deep diving, execute the TOC retrieval script: python "scripts/get_rag_toc.py" "keyword or title" (e.g. "Russell"). This returns a clean map of the book.
   • Execute the native retrieval script: python "scripts/search_rag.py" "your precise query".
   • Wait for the JSON response containing the retrieved contextual chunks.
2. Grade (Evaluate Relevance):
   • Critically analyze the returned text chunks. Do they actually contain the factual information needed to answer the user's question?
   • If a chunk is highly relevant, append its contents to your rag_search_state.
   • If the context is irrelevant or missing, discard it.
3. Adapt (Rewrite or Fallback):
   • If the first query failed or only partially answered the larger question, reformulate the query using different keywords, synonyms, or a broader concept.
   • Run the retrieval script again with the new query and repeat steps 1 & 2 to build the state.
   • If multiple local retrieval attempts fail to find the answer, you must fallback to a standard web search (e.g., using Tavily or direct web tools) to find the answer externally, appending that to the state.
4. Synthesize:
   • Once your rag_search_state contains sufficient information to fully address the user prompt, stop searching.
   • Ground your final answer strictly in the accumulated facts provided by the state.

Level 3 Resources

Scripts

• .agent/skills/retrieval/retrieving-rag-context/scripts/search_rag.py: Primary CLI utility to query the RAG via the active SSE endpoint.
  • Usage: python ".agent/skills/retrieval/retrieving-rag-context/scripts/search_rag.py" "your question"
  • Note: This streams raw JSON natively to standard out using UTF-8, making it safe and easy for you to parse.
• .agent/skills/retrieval/retrieving-rag-context/scripts/get_rag_toc.py: Secondary CLI utility to directly fetch the Table of Contents of a specific book.
  • Usage: python ".agent/skills/retrieval/retrieving-rag-context/scripts/get_rag_toc.py" "book title or keyword"
  • Note: This script calls OptimizedRAG.get_toc() directly — no MCP SSE server required. Falls back to MCP SSE if the direct import fails.

References

• references/agentic-logic.md: Breakdown of the Retrieve -> Grade -> Adapt decision tree.

Important Rules

• Maintain Accumulative State: Do not overwrite previous contextual discoveries. You are the orchestrator. You MUST accumulate relevant information across multiple searches before attempting to answer complex queries.
• Agentic Loop is Mandatory: You MUST grade the results yourself and adapt if the search fails. Do not blindly return bad context.
• Context Integrity: Respect the narrative context provided by the chunks. Do not hallucinate outside the given text unless explicitly falling back to web search.

Best Practices

• Parent Retrieval: Always retrieve parent chunks rather than small segments for better LLM grounding.
• Thresholding: The AgenticRAG system uses normalized keyword matching for relevance grading.
• Query Refinement: If results are poor, refine the query to be more specific to the domain terms found in the library.