Keppo

How it works

Four checks between a source and your agent.

Keppo extracts claims from official docs, declarations, issues, and verified examples. Your agent sees scoped answers with citations, freshness signals, and confidence labels.

  1. Step 1

    Extract

    The ingestion pipeline collects official docs, release notes, type declarations, and high-signal issue threads. Extractors turn them into atomic Behavior, Pattern, Compatibility, Safety, and Deprecation claims.

  2. Step 2

    Verify

    Where behavior can be executed, Keppo runs small check programs in isolated sandboxes and records the result as evidence. Docs-backed claims keep source snapshots and quote checks for currency.

  3. Step 3

    Synthesize

    Retrieval merges graph edges, FTS, and embeddings. The synthesizer answers only from retrieved evidence, then faithfulness checks demote unsupported prose into a research summary or abstention.

  4. Step 4

    Measure

    Answer tier, evidence grade, retrieval coverage, conflicts, stale risks, and streaming latency are captured so weak answers feed the enrichment queue instead of staying invisible.

For your agent

Two MCP tools, one loop.

Once Keppo is wired up over MCP, your agent can ask the public knowledge graph and report answer feedback directly from its tool loop.

  • ask_public_dev_knowledge

    Returns a cited, graded answer.

    ask_public_dev_knowledge({ query: "Why does Zod parse throw?" })
  • report_feedback

    Flags stale or incorrect answers.

    report_feedback({ question, answer_id, feedback_text })
Keppo