Cortex Pipeline

The Cortex is the shared LangGraph pipeline used by all Router agents. It processes user messages through a series of graph nodes, optionally retrieves context from Brain, and generates responses.

Pipeline Flow

router-cortex

Nodes

Each node in the LangGraph DAG represents an immutable state transition. By breaking the RAG process into distinct steps rather than a single massive LLM prompt, Cortex achieves higher determinism, allows for conditional loops, and produces transparent execution traces for the observability dashboard.

`extract_query`

Reads the latest HumanMessage from state, sets user_query, should_retrieve, and initializes defaults.

`detect_intent`

Uses a fast LLM (e.g., gemini-2.0-flash-lite) to classify the user’s intent:

Intent	Description	Uses RAG?
`rag_and_web`	Questions requiring knowledge retrieval	✅
`translate`	Translation requests	❌
`summarize`	Summarization requests	❌
`rewrite`	Rewriting/editing	❌
`identity`	Questions about the assistant itself	❌

Also performs taxonomy enrichment — matches user query against the taxonomy canonical_map to detect concepts and strengthen retrieval queries.

`retrieve`

Dual-path retrieval:

Path A — Vector Search:

Searches Vertex AI Search datastores (books, videos, QA, web)
Runs web search for allowed domains via Google CSE
Reranks results per source type using Vertex AI Ranking API

Path B — Graph Facts:

Fetches relationship facts from Brain Knowledge Graph
Filters by runtime_fact_labels from ontology configuration
Adds as non-citation background knowledge to the RAG prompt

`suggest`

Generates optional search suggestions for rag_and_web intent.

`generate`

Produces the final assistant response. If retrieval yielded zero documents, generates a no-results message.

Debugging

Enable pipeline debug logging:

export DEBUG_PIPELINE=1    # Per-node structured logs
export DEBUG_WEB_SEARCH=1  # Raw CSE result previews

Output format:

PIPELINE extract_query | user_query="How does RAG work?"
PIPELINE detect_intent.out | intent=rag_and_web taxonomy_concepts=["RAG"]
PIPELINE retrieve.out | docs=5 books=3 videos=1 qa=1 web=0
PIPELINE generate.out | assistant_chars=450 web_sources=0