Add LLM backend support and improve debugging observability

- Add LLM_BACKEND to environment configuration
- Implement detailed debug logging for LLM request/response cycles
- Add missing llama-index dependencies for embeddings and chroma
- Update prompt constraints to prevent lore redundancy
- Enable CUDA for transcription and set logging to DEBUG level
- Add entry point for running the orchestrator directly
- Cleanup unused comment in TUI context updates

.env

@@ -2,6 +2,7 @@
 OPENAI_API_KEY=no-key-required
 OPENAI_BASE_URL=https://vllm.tipsy.codes/v1
 LLM_MODEL=google/gemma-4-26b-a4b-it
+LLM_BACKEND=vllm
 #LLM_BACKEND=ollama
 #LLM_MODEL=gemma:2b
 WHISPER_MODEL=base

requirements.txt

@@ -9,3 +9,5 @@ python-dotenv
 llama-index
 chromadb
 pdfplumber
+llama-index-embeddings-huggingface
+llama-index-vector_stores-chroma

src/llm/processor.py

@@ -45,6 +45,7 @@ class LLMProcessor:
             final_base_url = base_url or os.environ.get("OPENAI_BASE_URL")
             final_api_key = api_key or os.environ.get("OPENAI_API_KEY")
 
+        logger.info(f"Using LLM backend: {backend}")
         try:
             self.client = OpenAI(
                 api_key=final_api_key,
@@ -96,6 +97,14 @@ class LLMProcessor:
 
         messages.append({"role": "user", "content": user_prompt})
 
+        # Debugging: Dump inputs
+        logger.debug("--- LLM CALL START ---")
+        logger.debug(f"Model: {self.model}")
+        logger.debug(f"Messages: {messages}")
+        if response_format:
+            logger.debug(f"Response Format: {response_format}")
+        logger.debug("--- LLM CALL END ---")
+
         try:
             response = self.client.chat.completions.create(
                 model=self.model,
@@ -105,6 +114,11 @@ class LLMProcessor:
             )
             content = response.choices[0].message.content
 
+            # Debugging: Dump outputs
+            logger.debug("--- LLM RESPONSE START ---")
+            logger.debug(f"Content: {content}")
+            logger.debug("--- LLM RESPONSE END ---")
+
             return self._strip_markdown_code_blocks(content)
         except Exception as e:
             logger.error(f"LLM Error: {e}")

src/llm/prompts.py

@@ -28,6 +28,7 @@ CONSTRAINTS:
 - OUTPUT ONLY VALID JSON.
 - If no relevant information is found, return empty lists for all keys.
 - If a character name is not specified (e.g., "Your character"), use "Player Character".
+- Do not repeat lore if it is already known; only provide new or updated facts.
 
 Strict Output Format:
 Return a JSON object with exactly these keys:

src/pipeline/orchestrator.py

@@ -23,7 +23,7 @@ from src.ui.tui import ConfirmationApp
 
 # Configure logging to write to a file instead of stdout
 logging.basicConfig(
-    level=logging.INFO,
+    level=logging.DEBUG,
     format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
     handlers=[
         logging.FileHandler("pipeline.log"),
@@ -44,7 +44,7 @@ class PipelineOrchestrator:
 
         # Modules
         self.listener = AudioListener(loop=self.loop)
-        self.transcriber = Transcriber(model_size="base")
+        self.transcriber = Transcriber(model_size="base", device="cuda")
         self.processor = LLMProcessor()
         self.rag_manager = RAGManager()
 
@@ -328,3 +328,16 @@ class PipelineOrchestrator:
         Stops.
         """
         self.is_running = False
+
+if __name__ == "__main__":
+    import asyncio
+
+    async def main():
+        loop = asyncio.get_event_loop()
+        orchestrator = PipelineOrchestrator(loop)
+        try:
+            await orchestrator.run()
+        except KeyboardInterrupt:
+            orchestrator.stop()
+
+    asyncio.run(main())

src/ui/tui.py

@@ -16,7 +16,7 @@ from textual.widgets import (
     Static,
 )
 
-from src.llm.models import CharacterStateUpdate, ExtractionResult, LoreUpdate
+from src.llm.models import CharacterStateUpdate, ContextUpdate, ExtractionResult, LoreUpdate
 from src.persistence.characters import update_character_state
 from src.persistence.lore import update_lore
 
@@ -218,8 +218,6 @@ class ConfirmationApp(App):
                 elif isinstance(update, ContextUpdate):
                     display_text = f"Query: {update.query}\nSource: {update.source}\n\n{update.snippet}"
                     context_list = self.query_one("#context-pane", ListView)
-                    # ListView.insert takes an *iterable* of ListItems; passing a
-                    # bare ListItem raises TypeError because ListItem is not iterable.
                     # Insert at the top to show most recent first.
                     await context_list.insert(0, [ListItem(Static(display_text))])
                 if hasattr(self.llm_to_ui_queue, "task_done"):