v2.3

New Features:

1. Compact Chat Memory
   • Added rolling conversation summary with the last two user/assistant exchanges.
   • Older raw messages are removed to prevent context growth.
   • Conversation memory is configurable through MAIN_HISTORY_MESSAGES_TO_KEEP.

2. Execution Logging
   • Added optional terminal logs for graph steps, routing decisions, tool calls, and outputs.
   • Added colored and truncated state previews for easier debugging.
   • Logging can be enabled from the project configuration.

3. Updated Model Defaults
   • Switched the default main LLM to granite4.1:8b.
   • Added deterministic generation with LLM_SEED = 42.
   • Updated the local judge model to ministral-3:3b-instruct-2512-q8_0.

Bug Fixes:

1. Prompt and Source Handling
   • Refined system prompts to reduce invented sources and generic answer headings.
   • Standardized source formatting across generator, fallback, and aggregation prompts.
   • Improved query rewriting for standalone terms, acronyms, and follow-up questions.

2. Retrieval and Chunking Cleanup
   • Restored the forced first retrieval call for local tool-calling reliability.
   • Changed parent chunk IDs from <document>_parent_N to <document>_pN.

3. Evaluation and Documentation Alignment
   • Improved the curated RAGAS QA dataset with balanced questions per document.
   • Updated evaluation to use a fresh chat thread for each query.
   • Synced prompts, README files, notebooks, Dockerfile, and project configuration.