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

@@ -1,7 +1,8 @@
 # D&D Helpers Configuration
 OPENAI_API_KEY=no-key-required
 OPENAI_BASE_URL=https://vllm.tipsy.codes/v1
-LLM_MODEL=Intel/gemma-4-31B-it-int4-AutoRound
+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/models.py

@@ -55,10 +55,6 @@ class ContextUpdate(BaseModel):
 
 
 class FilterResult(BaseModel):
-    contextual_info: str = Field(
-        ...,
-        description="Information interesting to the user but not useful for structured extraction",
-    )
     filtered_text: str = Field(
         ..., description="Cleaned transcript used for structured data extraction"
     )

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,
@@ -61,6 +62,18 @@ class LLMProcessor:
 
         self.model = model or os.environ.get("LLM_MODEL", "gpt-4o")
 
+    def _strip_markdown_code_blocks(self, content: str) -> str:
+        """
+        Strips markdown code blocks (e.g., ```json ... ```) from the content.
+        """
+        import re
+
+        # Remove opening and closing code blocks
+        content = re.sub(
+            r"^```(?:json)?\n?|```$", "", content, flags=re.MULTILINE
+        ).strip()
+        return content
+
     def _call_llm(
         self,
         system_prompt: str,
@@ -84,6 +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,
@@ -93,15 +114,12 @@ class LLMProcessor:
             )
             content = response.choices[0].message.content
 
-            # Strip markdown code blocks if present
+            # Debugging: Dump outputs
-            if content.startswith("```"):
+            logger.debug("--- LLM RESPONSE START ---")
-                import re
+            logger.debug(f"Content: {content}")
+            logger.debug("--- LLM RESPONSE END ---")
 
-                content = re.sub(
+            return self._strip_markdown_code_blocks(content)
-                    r"^```(?:json)?\n?|```$", "", content, flags=re.MULTILINE
-                ).strip()
-
-            return content
         except Exception as e:
             logger.error(f"LLM Error: {e}")
             return ""
@@ -147,34 +165,19 @@ class LLMProcessor:
         """
         logger.info(f"LLM Processor (Extract): Calling extraction for: {filtered_text}")
         try:
-            # Using standard chat.completions.create with JSON mode for better compatibility with vLLM
-            logger.info("LLM Processor (Extract): Sending request to backend...")
-
             system_prompt = EXTRACTION_SYSTEM_PROMPT
             if context:
                 system_prompt += f"\n{context}"
 
-            messages = [
+            result = self._call_llm(
-                {"role": "system", "content": system_prompt},
+                system_prompt=system_prompt,
-            ]
+                user_prompt=filtered_text,
-            messages.append({"role": "user", "content": filtered_text})
-
-            for message in messages:
-                logger.info(f"LLM Processor (Extract): Message: {message}")
-
-            response = self.client.chat.completions.create(
-                model=self.model,
-                messages=messages,
                 response_format={"type": "json_object"},
-                extra_body={"enable_thinking": False},
             )
-            logger.info("LLM Processor (Extract): Response received from backend.")
 
             import json
 
-            content = response.choices[0].message.content
+            data = json.loads(result)
-            logger.info(f"LLM Processor (Extract): Raw JSON response: {content}")
-            data = json.loads(content)
 
             # Map the JSON data to the Pydantic model
             return ExtractionResult(**data)

src/llm/prompts.py

@@ -12,7 +12,6 @@ NOISE_FILTER_SYSTEM_PROMPT = """
 You are a D&D Game Master's assistant. Given a transcript, remove all out-of-character (OOC) chatter, logistical discussions (e.g., 'Where is my d20?'), and non-relevant noise.
 
 You must output your response as a JSON object with the following keys:
-- "contextual_info": Information that is interesting or relevant to the story/session but doesn't fit into lore, character state, or significant events (e.g., flavor text, atmospheric descriptions, player commentary that adds context).
 - "filtered_text": The cleaned transcript. IMPORTANT: Keep all player questions, requests for rule clarifications, and mentions of spells, NPCs, or locations in this field, as they are used to trigger knowledge base lookups.
 
 Keep the original speakers' names if they are present in the transcript.
@@ -22,15 +21,14 @@ Do not add any commentary or summaries. Just filter the text.
 EXTRACTION_SYSTEM_PROMPT = """
 You are a D&D session analyzer. Your goal is to extract structured data from a filtered transcript.
 Extract any changes to character states (HP, status effects, inventory) and any new lore facts (NPCs, locations, world-building).
-
+In addition extracting updates to character state and lore, look for the oppertunity to provide useful context,
-DO NOT THINK.
+such as the answer to a player's question or the resolution of a lore fact.
 
 CONSTRAINTS:
 - OUTPUT ONLY VALID JSON.
-- DO NOT include any commentary, explanations, or "thought" blocks.
-- DO NOT include any keys other than "lore", "character_state", and "events".
 - 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="small")
+        self.transcriber = Transcriber(model_size="base", device="cuda")
         self.processor = LLMProcessor()
         self.rag_manager = RAGManager()
 
@@ -54,6 +54,7 @@ class PipelineOrchestrator:
         self.clean_to_llm_queue = asyncio.Queue()
         self.llm_to_ui_queue = asyncio.Queue()
         self.log_queue = asyncio.Queue()
+        self.persistence_queue = asyncio.Queue()
 
         self.is_running = False
 
@@ -107,7 +108,9 @@ class PipelineOrchestrator:
                 full_audio = np.concatenate(self.audio_buffer)
 
                 # Transcribe (WhisperX now returns a list of (speaker, text, start, end))
-                results = self.transcriber.transcribe(full_audio)
+                results = await asyncio.to_thread(
+                    self.transcriber.transcribe, full_audio
+                )
 
                 # Filter for only new segments that start after the last processed segment
                 new_segments = [
@@ -184,8 +187,11 @@ class PipelineOrchestrator:
         async def feed_ui():
             while self.is_running:
                 try:
-                    text = await self.ui_to_llm_queue.get()
+                    item = await self.ui_to_llm_queue.get()
-                    await internal_queue.put(("UI", text))
+                    if isinstance(item, (LoreUpdate, CharacterStateUpdate)):
+                        await self.persistence_queue.put(item)
+                    else:
+                        await internal_queue.put(("UI", item))
                 except Exception as e:
                     logger.error(f"LLM Feeder (UI) error: {e}")
 
@@ -213,20 +219,8 @@ class PipelineOrchestrator:
                     context=context,
                 )
 
-                # Persistence: Lore Updates
+                # Send the entire result to UI for confirmation
-                for lore_update in extraction_result.lore_updates:
+                await self.llm_to_ui_queue.put(extraction_result)
-                    file_path = await asyncio.to_thread(update_lore, lore_update)
-                    await asyncio.to_thread(self.rag_manager.ingest_file, file_path)
-                    logger.info(
-                        f"LLM Worker: Lore updated and ingested into RAG: {lore_update.entity_name}"
-                    )
-
-                # Persistence: Character State Updates
-                for char_update in extraction_result.character_updates:
-                    await asyncio.to_thread(update_character_state, char_update)
-                    logger.info(
-                        f"LLM Worker: Character {char_update.character_name} state updated."
-                    )
 
                 # UI Notification: Context Updates
                 for context_update in extraction_result.context_updates:
@@ -243,29 +237,32 @@ class PipelineOrchestrator:
         for f in feeders:
             f.cancel()
 
-    def _get_wiki_context(self) -> str:
+    async def persistence_worker(self):
         """
-        Reads all files in the lore directory and returns them as a 저희 context string.
+        Worker that handles persistence: Confirmed updates -> Disk & RAG.
         """
-        from src.persistence.lore import DATA_LORE_DIR
+        logger.info("Persistence Worker started.")
-
+        while self.is_running:
-        wiki_contents = []
-        # Recursively find all .md files in the lore directory
-        for path in DATA_LORE_DIR.rglob("*.md"):
             try:
-                with open(path, "r", encoding="utf-8") as f:
+                update = await self.persistence_queue.get()
-                    content = f.read()
+                if isinstance(update, LoreUpdate):
-                    wiki_contents.append(
+                    file_path = await asyncio.to_thread(update_lore, update)
-                        f"File: {path.relative_to(DATA_LORE_DIR)}\nContent:\n{content}"
+                    await asyncio.to_thread(self.rag_manager.ingest_file, file_path)
+                    logger.info(
+                        f"Persistence Worker: Lore updated and ingested into RAG: {update.entity_name}"
+                    )
+                elif isinstance(update, CharacterStateUpdate):
+                    await asyncio.to_thread(update_character_state, update)
+                    logger.info(
+                        f"Persistence Worker: Character {update.character_name} state updated."
                     )
-            except Exception as e:
-                logger.error(f"Error reading wiki file {path}: {e}")
 
-        return (
+                if hasattr(self.persistence_queue, "task_done"):
-            "\n\n".join(wiki_contents)
+                    self.persistence_queue.task_done()
-            if wiki_contents
+            except Exception as e:
-            else "No wiki knowledge available."
+                logger.error(f"Persistence Worker error: {e}")
-        )
+
+            await asyncio.sleep(0.1)
 
     async def tui_worker(self):
         """
@@ -306,6 +303,7 @@ class PipelineOrchestrator:
             asyncio.create_task(self.stt_worker()),
             asyncio.create_task(self.clean_worker()),
             asyncio.create_task(self.llm_worker()),
+            asyncio.create_task(self.persistence_worker()),
             asyncio.create_task(self.tui_worker()),
         ]
 
@@ -330,3 +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
 
@@ -213,12 +213,13 @@ class ConfirmationApp(App):
         while True:
             try:
                 update = await self.llm_to_ui_queue.get()
-                display_text = f"Query: {update.query}\nSource: {update.source}\n\n{update.snippet}"
+                if isinstance(update, ExtractionResult):
-                context_list = self.query_one("#context-pane", ListView)
+                    self.handle_proposal_result(update)
-                # ListView.insert takes an *iterable* of ListItems; passing a
+                elif isinstance(update, ContextUpdate):
-                # bare ListItem raises TypeError because ListItem is not iterable.
+                    display_text = f"Query: {update.query}\nSource: {update.source}\n\n{update.snippet}"
-                # Insert at the top to show most recent first.
+                    context_list = self.query_one("#context-pane", ListView)
-                await context_list.insert(0, [ListItem(Static(display_text))])
+                    # 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"):
                     self.llm_to_ui_queue.task_done()
             except Exception as e:
@@ -263,17 +264,15 @@ class ConfirmationApp(App):
             self.ui_to_llm_queue.put_nowait(text)
             input_widget.value = ""
 
-    def action_accept(self) -> None:
+    async def action_accept(self) -> None:
         table = self.query_one("#pending-facts-table", DataTable)
         row_index = table.cursor_row
         if row_index < 0 or row_index >= len(self.pending_updates):
             return
 
         update = self.pending_updates[row_index]
-        if isinstance(update, LoreUpdate):
+        if self.ui_to_llm_queue:
-            update_lore(update)
+            self.ui_to_llm_queue.put_nowait(update)
-        elif isinstance(update, CharacterStateUpdate):
-            update_character_state(update)
 
         self.remove_update(row_index)