Session memory: condense-then-retrieve

[markgewhite]

Summary

• Added Condenser module that rewrites follow-up questions into standalone queries via Ollama LLM call
• First/independent questions (no chat history) skip condensation — no LLM overhead
• Chat history tracked per session, kept separate from answer generation context window
• Standalone question used for both retrieval and reranking

Closes #9

Test plan

• Condenser skips condensation when no chat history
• Condenser rewrites follow-up questions using chat history
• Correct model passed to LLM call
• History included in condensation prompt
• Output whitespace stripped
• Integrated into Chainlit pipeline
• Full test suite: 82 tests pass

🤖 Generated with Claude Code