Nanocortex Web Agent

A knowledge agent that surfaces connected answers, not fragments

Zettelkasten retrieval, human approval and admin control, built on the Snowflake Cortex Agent API.

Kevin Keller · kevinkeller.org

Why it matters

Staff need consistent answers from large document collections

  • Insurance policies, legal contracts, IT runbooks, HR procedures.
  • AI-generated content in regulated industries needs human sign-off.
  • Content owners need to know which knowledge is used and which is wrong.
  • Operations teams want to tune the agent without a deployment.

From CLI to application

The CLI proved the loop. The web agent proves the application.

  • Follows Nanocortex, a minimal single-file CLI for the Cortex Agent API.
  • The demo is an Acme Insurance agent: policy Q&A, claims, underwriting.
  • Skills, knowledge files, topics and prompts are swappable without code changes.

The idea

Knowledge is a graph, not a pile of chunks

Ask about a deductible and the agent also surfaces coinsurance, out-of-pocket maximums and the claims workflow, because the graph connects them.

How retrieval ranks

Well-connected, important knowledge surfaces first

  • Each knowledge entry becomes a node in a semantic graph.
  • Explicit connects-to links define relationships.
  • Semantic similarity auto-links related entries.
  • Ranking combines similarity, importance, recency and graph connectivity.

Skills versus knowledge

Separate how the agent acts from what it knows

SkillsKnowledge
PurposeHow to actWhat to know
Format.md in skills/.knowledge.md in knowledge/
DeliveryMandatory, keyword-triggered or user-selectedGraph search, injected at query time

Human in the loop

Every substantive answer becomes a reviewable document

  1. Draft: the agent creates an editable workspace document.
  2. In Review: reviewers are assigned and comments added.
  3. Approved: content circulates only after expert sign-off.

Versions are tracked throughout.

Operational control

Tune the agent from a dashboard, not a deployment

  • Model: switch Sonnet or Opus, set thinking effort
  • Behaviour: edit system prompt and skills
  • Tools: toggle web search per deployment
  • Knowledge: ingest, re-index or remove files
  • Users: roles, permissions, session history

Feedback loop

Analytics show where the knowledge base needs work

  • Asset reuse: high reuse signals content worth expanding.
  • Review flags: flagged answers land in a review queue.
  • Topic distribution and response times per session.
  • A self-review reflection pass corrects answers before delivery.

Deployment choices

Snowflake for one platform, PostgreSQL for local work

SnowflakePostgreSQL + pgvector
Best forProductionLocal and self-hosted
StorageEverything in SnowflakeLocal
InferenceCortexCortex

Production sign-in uses magic links and WebAuthn passkeys.

The takeaway

Swap the knowledge files and the same graph, workflow and controls serve any domain.

Read the full article

Kevin Keller · Open-source project on GitHub