Claude Chat CLI with Hybrid RAG Retrieval

Build and inspect a multi-turn Claude CLI with streaming responses, native web search, hybrid local retrieval, source grounding, and prompt evaluation.

Get started Explore the architecture

The deterministic demo follows the same prompts, streaming behavior, and local-source format as the real CLI.

What the repository demonstrates

• A streaming Anthropic Messages API client with persistent conversation history.
• Optional extended thinking, prompt caching, and native web search.
• AST-aware Python and heading-aware Markdown chunking.
• Okapi BM25 and exact cosine vector search implemented without a search framework.
• Reciprocal Rank Fusion across lexical and semantic rankings.
• Grounded prompts with citations, abstention behavior, and prompt-injection boundaries.
• Deterministic retrieval evaluation and LLM-as-judge prompt evaluation.

Choose a path

Goal	Start here
Run the finished assistant	Get started
Understand retrieval and generation boundaries	Architecture
Inspect the implementation	Repository on GitHub

Guide automation is review-gated

Executable tutorial commands are verified in CI. The generation workflow can draft one planned offline tutorial as a pull request, but facts, curriculum, rendering, and human editorial review must all pass before it can merge.