retrieving-rag-context-adv

Patterns for using the antigravity-rag tools to ingest documents into Qdrant and perform semantic retrieval (Parent-Child strategy).

When to Use

• When the assistant needs to answer questions based on documents that are not part of its base training data.
• To provide source-grounded answers for technical queries.
• When new domain-specific knowledge is added to the project via external files.

Prerequisites

• antigravity-rag server must be running and configured in mcp_config.json.
• Documents must be in supported PDF format (standardized for Parent-Child).
• Python environment with qdrant-client and fastembed installed.

Tools Used

• mcp_antigravity-rag_ingest_document: Adds external documents (PDF) to the Qdrant vector store.
• mcp_antigravity-rag_search_library: Performs semantic search and retrieves parent chunks for rich context.
• mcp_antigravity-rag_list_library_sources: Lists all unique documents currently indexed.
• mcp_local-faiss-mcp_search_semantic: Performs pure semantic search without LLM generation.

Process

1. Document Ingestion

   • Use ingest_document to index new knowledge.
   • Example: mcp_local-faiss-mcp_ingest_document(path="./docs/manual.pdf")

2. RAG Workflow

   • When a user asks a question about indexed content, use query_rag_store.
   • The tool will perform the retrieval and return the synthesized answer with sources.
   • Example: mcp_local-faiss-mcp_query_rag_store(query="How do I configure the server?")

3. Pure Semantic Search

   • If you only need relevant chunks without an LLM-synthesized answer, use search_semantic.
   • Example: mcp_local-faiss-mcp_search_semantic(query="installation steps")

Patterns

Handling New Data

When the user provides new files or links that should be "remembered":

1. Save the file to a local path if it isn't already.
2. Call ingest_document.
3. Confirm ingestion to the user.

Question Answering

When the user asks "How do I..." or "What is..." regarding the project context:

1. Call search_library first to see if the answer is in the vector base.
2. If no relevant info is found, fall back to general knowledge or other retrieval tools.

Best Practices

• Metadata first: Ensure documents have clear titles and versions before ingestion.
• Query refinement: If retrieval fails, try rephrasing the query to use more specific keywords.
• Chunk size: Use the default 1000-character chunk size for balanced retrieval accuracy.
• Verify ingestion: Always check the catalog after ingestion to ensure the document was correctly processed.