Knowledge Base Signal

Overview

kb binds routing signals to the output of a named knowledge base instance. Define these bindings under routing.signals.kb.

Use it for an embedding-backed knowledge base referenced by a reachable recipe. Only referenced KBs are prepared at startup; an unused catalog entry does not load its assets or start an embedding model.

Key Advantages

• Reuses one exemplar set across several routes.
• Keeps labels, groups, and numeric metrics explicit instead of relying on magic runtime names.
• Supports both winner-style and threshold-style signal bindings.
• Lets projections consume continuous knowledge base metrics without turning signals into a scripting surface.

What Problem Does It Solve?

Some routing policies depend on a curated exemplar set rather than a single keyword or embedding candidate list. One knowledge base can classify requests into privacy, safety, emotion, or preference labels, while the routing config exposes only the labels or groups that decisions need.

kb keeps that split explicit:

• global.model_catalog.kbs[] owns the reusable knowledge base package
• routing.signals.kb[] binds specific labels or groups into named routing signals
• routing.projections can consume knowledge base metrics such as best_score, best_matched_score, or configured group margins

When to Use

Use kb when:

• requests must be classified against a curated exemplar set
• one startup-loaded knowledge base result should feed several routes
• you want stable route-level groups without duplicating exemplars
• you need explicit bindings instead of implicit signal names

Configuration

global:
  model_catalog:
    kbs:
      - name: privacy_kb
        source:
          path: knowledge_bases/privacy/
          manifest: labels.json
        threshold: 0.55
        prototype_scoring:
          enabled: true
          cluster_similarity_threshold: 0.9
          max_prototypes: 8
          best_weight: 0.75
          top_m: 2
          margin_threshold: 0.05
        label_thresholds:
          prompt_injection: 0.7
        groups:
          privacy_policy: [proprietary_code, internal_document, pii]
          security_containment: [prompt_injection, credential_exfiltration]
          private: [proprietary_code, internal_document, pii, prompt_injection, credential_exfiltration]
          public: [generic_coding, general_knowledge]
        metrics:
          - name: private_vs_public
            type: group_margin
            positive_group: private
            negative_group: public

routing:
  signals:
    kb:
      - name: privacy_policy
        kb: privacy_kb
        target:
          kind: group
          value: privacy_policy
        match: best
      - name: proprietary_code
        kb: privacy_kb
        target:
          kind: label
          value: proprietary_code
        match: threshold

Keep knowledge base names stable because kb signals bind to those names directly.

When prototype_scoring is enabled, the KB builds per-label prototype banks from the label exemplars. Runtime classification then scores labels from those label-owned prototypes instead of letting one raw exemplar dominate the whole label forever.

Match Semantics

routing.signals.kb[] supports:

• target.kind: label or group
• match: best or threshold

Meaning:

• label + best: match only when the label is the knowledge base's best label
• label + threshold: match when the label score clears its effective threshold
• group + best: match only when the group is the knowledge base's best group
• group + threshold: match when any member label clears its threshold

Projection Metrics

Knowledge base signals are boolean routing inputs. Numeric outputs stay in routing.projections.

For example:

routing:
  projections:
    scores:
      - name: privacy_bias
        method: weighted_sum
        inputs:
          - type: kb_metric
            kb: privacy_kb
            metric: private_vs_public
            value_source: score
            weight: 1.0

Named knowledge base metrics are declared under global.model_catalog.kbs[].metrics[]. Built-in metrics best_score and best_matched_score are always available.

Dependencies and Limitations

• The knowledge-base package is loaded from global.model_catalog.kbs[].source. Keep its manifest and files versioned together.
• Request text is embedded through the shared semantic embedding runtime. A remote embedding provider therefore receives the text.
• Labels, groups, thresholds, and embedding model form one calibrated unit; re-evaluate them together when any part changes.
• See the knowledge-base signal example config/fragments/signal/kb/privacy.yaml and the complete KB declaration in config/config.yaml.