Skip to content
Factflow

Knowledge workflow

Knowledge is three subpackages of adapters covering the lifecycle of a concept map:

  • Detection — identify concepts in segmented content
  • Consolidation — merge across sources; write to Avalon (DNB’s knowledge platform)
  • Diff — compare two snapshots; emit a structured diff

Each stage is a pipeline adapter. Compose them into a route chain or run them separately.

Typical pipeline

Section titled “Typical pipeline”
version: "1.0"


routes:
  concept_detection:
    inbound:
      queue: "/queue/knowledge.concepts.in"
      subscription: "concept-detectors"
      concurrency: 2   # LLM-bound
    adapters:
      - type: "concept_detection"
        config:
          provider: "default"
          model: "claude-sonnet-4-6"


  concept_consolidation:
    inbound:
      queue: "/queue/knowledge.consolidation.in"
      subscription: "consolidators"
    adapters:
      - type: "concept_consolidation"
        config:
          avalon_endpoint: "${AVALON_ENDPOINT}"


  knowledge_diff:
    inbound:
      queue: "/queue/knowledge.diff.in"
      subscription: "differs"
    adapters:
      - type: "knowledge_diff"
        config:
          diff_storage_prefix: "diffs/"

Export CLI

Section titled “Export CLI”

Render a concept map to a shareable format:

Terminal window
cd backend
uv run python -m factflow_knowledge.export \
  --origin "boost:FIX" \
  --format tree \
  --output ./concepts-fix.html

Formats: tree, report, taxonomy, stats. See factflow_knowledge.export.renderer for extension points.

Avalon integration

Section titled “Avalon integration”

The concept_consolidation adapter writes to Avalon via HTTP. Credentials via env:

AVALON_ENDPOINT=https://avalon.dnb.internal
AVALON_API_KEY=...

Testing epistemic quality

Section titled “Testing epistemic quality”

The epistemic-evaluation skill in .claude/skills/ captures patterns for assessing extraction quality (precision, recall, coherence) against reference datasets.

Section titled “Related”