Skip to main content

Confidence = trust × relevance

The confidence score multiplies retrieval relevance by a trust score built from where the content came from, how recently it was written, and whether it matches the version you target.

The formula is explicit. Every hit carries its per-factor breakdown, so your AI assistant can explain why a passage ranked where it did without a second round-trip.

The trust factors​

Trust is a composite of five independent signals:

Attribution tier​

Who published the content? Five tiers, each weighted differently:

  • Foundation: produced directly by the Midnight Network Foundation; highest trust.
  • Partner: produced by an accredited ecosystem partner.
  • Third-party: independent developer or community project.
  • Community: informal or crowdsourced contribution.
  • Unknown: provenance could not be established.

Foundation-authored content starts with a significant advantage; unknown-provenance content starts penalized.

Verification status​

Has a human reviewed and vouched for the content? Verified content earns a boost; unverified content does not. The verification chain records who did the verifying (Foundation, partner, or community member), so the boost is proportional to the verifier's authority.

Freshness​

Documentation that was accurate six months ago may be wrong today. Midnight Manual applies exponential decay by age: a passage written last week scores materially higher than an equivalent one from eighteen months ago, all else equal. Fast-moving docs (SDK changelogs, compiler release notes) don't sneak to the top by sheer volume of old references.

Deprecation flag​

Content explicitly marked deprecated is down-weighted rather than hidden. It may still surface as a last resort, but the confidence score reflects its status and the factor breakdown will name it. Your assistant can tell a user "this is the old API; here is the current one."

Version match​

Version match is the only factor that can exclude a chunk outright; the others only adjust its weight. The corpus tracks which Compact language version, SDK version, or component version each chunk belongs to. At query time:

  • Content that satisfies your target version is boosted.
  • A near-miss (adjacent version) is penalized in proportion to how far off it is.
  • A breaking mismatch is excluded entirely.
  • In strict mode, only version-satisfying content passes at all.

Version targets are extracted automatically at ingest (from pragma language_version in Compact files and from package.json / Cargo.toml manifests), so the corpus carries version metadata without manual tagging.

The factor breakdown​

Nothing about the confidence score is hidden. Each result carries the factor breakdown, so:

  • An assistant can say "this is Foundation-authored, recently verified, and matches the SDK version you specified" rather than just "here is a result."
  • A downstream tool can filter on individual trust signals (showing only verified content for a security-sensitive query, for example).
  • You can tune the weights without a rebuild: the scoring policy is loaded from a data file at runtime.

How it interacts with retrieval​

Confidence does not replace the retrieval score; it multiplies it. A passage with high semantic relevance but zero trust (say, unverified community content from two years ago on a deprecated API) ranks behind a passage that is moderately relevant but Foundation-authored, recent, and version-matched.

The reranker (rerank-2.5) sees the original candidates; the confidence multiplier is applied after reranking to produce the final order. This keeps the two signals orthogonal: the reranker optimizes for semantic fit; confidence adjusts for provenance.

  • Models — the reranker that sharpens the candidate set before confidence scaling.
  • Hybrid retrieval & RRF — how the candidate set is built in the first place.
  • Multi-query / HyDE — techniques for improving recall before confidence scoring.