Move contextual information handling from noise filtering to extraction
and centralize LLM call logic. Wrap blocking transcription and state
update calls in asyncio.to_thread to prevent event loop blocking.
Update transcriber model size to base.
- Add log_queue to PipelineOrchestrator and log LLM inputs to UI
- Use entity_name for lore update logs instead of topic
- Pass log_queue into ConfirmationApp to display logs in UI
- Introduce a log pane and left/right pane layout in the UI
- Poll and render log messages via a new poll_log_updates worker
- Run log polling with Textual workers to avoid GC issues
- Fix ListView insertion by wrapping ListItem in a list
- Relax RAG similarity threshold from 0.7 to 0.5
- Add ContextPipeline for async RAG lookups
- Implement RAG result summarization via LLMProcessor
- Add CLI flag for PDF ingestion
- Strip markdown code blocks from LLM responses
- Update TUI context display to use ListItems
- Add RAG capabilities using LlamaIndex and ChromaDB
- Implement RAGManager for PHB indexing and retrieval
- Integrate RAG pipeline into orchestrator to trigger queries based on extracted entities
- Update TUI to include a 3-column layout with a real-time Context Pane
- Define ContextUpdate data models in src/llm/models.py
- Update requirements.txt with new dependencies
Implement a sliding window audio buffer and update the transcriber to
use WhisperX for transcription, alignment, and speaker identification.
Update the pipeline to handle and store speaker-attributed transcripts.
Additionally, update the LLM processor's reasoning parameter to
"enable_thinking".
- Implement Silero VAD for dynamic audio chunking
- Add support for Ollama and vLLM backends
- Harden extraction prompts for strict JSON output
- Refactor TUI worker to handle proposals asynchronously
Add AGENTS.md to document the delegated agent architecture. Introduce
the Validator persona and update the Codemonkey and Orchestrator
personas to improve quality assurance and communication.
charles