Search quality

Seekite combines typo-tolerant lexical retrieval with optional semantic ranking. Quality work is treated as an evidence problem: keep a representative query set, record relevance judgments, and compare every ranking change to an accepted baseline.

Tier 1 behavior

The format-v2 runtime provides the features that most documentation and static content search needs without loading every shard up front:

• Unicode-aware token folding and optional English stemming;
• prefix expansion for the final query token and bounded fuzzy matching, so a query such as typo tolarance still finds this page;
• BM25F field weights for title, heading, and body text;
• filters and facet counts over declared metadata columns;
• grouping by document, pagination, and lazily hydrated content snippets; and
• hybrid scoring when an embedding provider matching the index manifest is available, with a lexical-only fallback when it is not.

Fuzzy expansion is deliberately bounded. Short terms are not expanded into a large neighborhood, candidates come from the lexical dictionary, and exact or prefix matches remain more valuable than edit-distance matches.

Configure the corpus

Declare searchable metadata at build time so the portable corpus contains only the columns the UI needs:

import { defineSearch, staticVectors } from "seekite"
import { seekiteEmbeddings } from "@seekite/embeddings-ternlight"

export default defineSearch({
  corpora: {
    docs: {
      source: docs,
      vectors: staticVectors(seekiteEmbeddings()),
      lang: "en",
      filters: ["section", "version"],
      facets: ["section"],
      fields: { title: 2.4, heading: 1.5, body: 1 },
    },
  },
})

The query API controls ranking and result shape without changing the index:

const response = await search.query("typo tolarance", {
  corpora: ["docs"],
  mode: "hybrid",
  semanticWeight: 0.35,
  where: { section: "Guides" },
  facets: ["section"],
  group: "document",
  hydrate: true,
  limit: 10,
})

Use mode: "lexical" when deterministic text matching is the entire job, or mode: "semantic" to inspect the vector ranker in isolation. Hybrid mode is the normal product default, not a substitute for measuring both components.

Measure a change

Put hand-reviewed relevance judgments in the query fixture configured under bench.queries, then create a run file:

seekite bench --label before-ranking-change
seekite bench --baseline .seekite/bench/before-ranking-change.json

Review Recall@k, MRR, and NDCG for lexical and hybrid candidates together with query latency and artifact size. Synthetic queries are useful retrievability checks, but promote important cases to human judgments before treating their scores as relevance evidence. Attach the baseline delta to every pull request that changes analysis, chunking, embeddings, quantization, or ranking.