rag-pipeline
Details on the Retrieval Augmented Generation pipeline, Ingestion, and Vector Search.
When & Why to Use This Skill
This Claude skill provides a comprehensive Retrieval Augmented Generation (RAG) pipeline, enabling agents to handle complex data ingestion, vector-based search, and sophisticated context management. By integrating Qdrant for vector storage and utilizing advanced embedding models, it allows for the creation of highly context-aware agents that can retrieve and process information from local documentation with high precision.
Use Cases
- Technical Documentation Assistant: Automatically ingest and index MDX-based documentation to provide precise, context-rich answers to complex technical queries.
- Personalized AI Interaction: Leverage user-specific software and hardware backgrounds to tailor the RAG prompt, ensuring the agent's responses are optimized for the user's specific environment.
- Automated Knowledge Base Management: Streamline the workflow of scanning, chunking, and embedding new project files into a vector database for real-time information retrieval.
- Agentic Workflow Integration: Use the custom Agent class to wrap LLM calls with retrieved context, facilitating the development of multi-agent systems that share a unified knowledge source.
| name | RAG Pipeline |
|---|---|
| description | Details on the Retrieval Augmented Generation pipeline, Ingestion, and Vector Search. |
RAG Pipeline Logic
Ingestion
- Script:
backend/ingest.py - Process:
- Scans
docs/. - Cleans MDX (removes frontmatter/imports).
- Chunks text (1000 chars, 100 overlap).
- Embeds using
models/text-embedding-004. - Upserts to Qdrant collection
physical_ai_book.
- Scans
- Run:
python backend/ingest.py
Vector Search (Qdrant)
- Client:
qdrant-client - Collection:
physical_ai_book - Vector Size: 768 (Gecko-004)
- Similarity: Cosine
Prompt Engineering
- File:
backend/utils/helpers.py. - RAG Prompt: Constructs a prompt containing retrieved context chunks.
- Personalization:
backend/personalization.pycreates system instructions based onsoftware_backgroundandhardware_backgroundof the user.
Agentic Flow
We use a custom Agent class (backend/agents.py) that wraps the LLM calls, allowing for future expansion into multi-agent workflows.