Kern

RAG

Hybrid search grounding text-to-SQL in the Dash team.

The Dash team uses retrieval to ground every SQL query in known patterns. Hallucinated SQL is the failure mode RAG prevents.

What it retrievesWhy retrieval
Table descriptions, validated SQL queries, business rulesThe Analyst grounds every query in known patterns. Without retrieval, the model invents column names and misreads enum values.

Dash: SQL grounded in known patterns

Dash holds three kinds of knowledge under one Knowledge instance:

KnowledgePurpose
Table metadataColumn meanings, value enums, gotchas
Validated query patternsTested SQL the Analyst can adapt
Business rules"MRR excludes trials", "active means ended_at IS NULL"

When a user asks "what's our MRR trend?", the Analyst retrieves matching table descriptions, similar past queries, and the MRR definition. The model writes SQL with all of that in context. Hallucination rate drops.

1from kern.knowledge import Knowledge
2from kern.vector_db.pgvector import PgVector
3
4dash_knowledge = Knowledge(
5    vector_db=PgVector(
6        table_name="dash_knowledge",
7        db_url=DB_URL,
8        search_type="hybrid",  # semantic + keyword
9    ),
10)
11
12analyst = Agent(
13    knowledge=dash_knowledge,
14    add_knowledge_to_context=True,
15    search_knowledge=True,
16)

search_type="hybrid" runs both vector similarity and BM25 keyword search, then merges results. Catches both "what's the same idea worded differently?" (semantic) and "find the doc that mentions this exact term" (keyword).

Loading knowledge

Knowledge content gets loaded once at boot or via a script:

1# Dash: table metadata, queries, business rules
2python -m agents.dash.scripts.load_knowledge

Re-running with --recreate rebuilds from scratch. Without the flag, content gets upserted by primary key.

How retrieval gets injected

When add_knowledge_to_context=True, AgentOS:

  1. Takes the user's message.
  2. Runs hybrid search against the Knowledge vector DB.
  3. Pulls top-k chunks with metadata (configurable).
  4. Injects them into the system prompt under a "Relevant context" section.
  5. The model answers with both the message and the retrieved context.

If search_knowledge=True is also set, the agent gets a search_knowledge_base(query) tool and can run additional searches mid-run when it needs to.

See it in action

Try in chatWhat happens
Dash: "what's our MRR trend?"Analyst retrieves MRR definition + matching query, writes grounded SQL.
Dash: "why is churn so high?"Analyst retrieves churn definitions, finds matching query patterns, executes, interprets.
Dash: "show me revenue by plan"Analyst pulls table metadata for subscriptions, generates the aggregation.

When hybrid search isn't enough

For very large knowledge bases, add reranking. PgVector + a reranker (Cohere, BGE) tightens the top-k:

1from kern.vector_db.pgvector import PgVector
2from kern.rerank.cohere import CohereReranker
3
4dash_knowledge = Knowledge(
5    vector_db=PgVector(
6        table_name="dash_knowledge",
7        db_url=DB_URL,
8        search_type="hybrid",
9        reranker=CohereReranker(model="rerank-3.5"),
10    ),
11)

Two-stage retrieval (hybrid then rerank) is the standard production setup. The hybrid stage casts a wide net (top-50), the reranker prunes to the best top-10.

Source: agents/dash/

Next

Knowledge →