LLM Pipeline

LLM Pipeline Skill

2. Scoring (AI Judge, not heuristics - ADR-003)

score = await importance_scorer.score(message)

classification: SIGNAL (>0.6) / NOISE (<0.3)

3. Auto-trigger extraction when threshold met

if unprocessed_count >= 10: # ai_config.message_threshold await extract_knowledge_from_messages_task.kiq()

4. KnowledgeOrchestrator runs Pydantic AI agent

agent = Agent( model=model, system_prompt=get_extraction_prompt("uk"), output_type=KnowledgeExtractionOutput, # CRITICAL: structured output output_retries=5, ) result = await agent.run(messages_content)

5. Save to DB + embed

await save_topics_and_atoms(result.output) await embed_atoms_batch_task.kiq(atom_ids)

</extraction-flow>

<agent-creation>
```python
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIChatModel

# Provider-specific model creation
if provider.type == "ollama":
    model = OpenAIChatModel(
        model_name=agent_config.model_name,
        provider=OllamaProvider(base_url=provider.base_url),
    )
elif provider.type == "openai":
    model = OpenAIChatModel(
        model_name=agent_config.model_name,
        provider=OpenAIProvider(api_key=api_key),
    )

# Agent with structured output
agent = Agent(
    model=model,
    output_type=MyPydanticModel,  # Forces JSON schema
    system_prompt="...",
    output_retries=5,
)

await embedding_service.generate_embedding(text) await embedding_service.embed_messages_batch(session, ids, batch_size=10)

</embedding-service>

<rag-context>
```python
# SemanticSearchService uses pgvector cosine similarity
similar_atoms = await search_service.search_atoms(
    query_embedding=embedding,
    limit=5,
    threshold=0.65,  # ai_config.semantic_search
)

# RAGContextBuilder assembles context for LLM
context = await rag_builder.build_context(
    query=user_query,
    similar_atoms=similar_atoms,
    related_messages=messages,
)

Hybrid: Project context (CAG) + similar atoms (RAG) = best extraction quality. See: @references/rag.md for detailed comparison.