ECAI Search: Find the Record and Check the Evidence

When an assistant quotes a document, a reader should be able to follow the reference. Which record was retrieved? Which collection contained it? Can another service check that the supplied record belongs to the published index?

ECAI has building blocks for answering those questions. Its live search path returns source records with scores and previews. Its artifact and proof paths can bind a record to a particular index release. Together, they offer a useful foundation for assistants that show their supporting material and for services that exchange independently checkable corpus references.

Applications still have to connect those pieces. A normal search response does not automatically attach a version 2 membership proof to every result, and a retrieved passage still needs interpretation in context.

The shared term vocabulary extracts keys from fields such as title, name, category, city, phone, text, abstract, language and Wikidata identifier. It includes exact terms and selected prefixes, suffixes and phone fragments. Text and abstract token limits bound how much of a record contributes terms.

The live search engine gathers the corresponding posting lists and scores their documents. Field weights favour some kinds of match over others; term rarity and multiple matching terms affect the score. The code also includes dataset-specific signals such as business review counts and Wikimedia visibility data.

This is lexical retrieval with explicit scoring rules. It can be useful for named entities, titles, directory records and source excerpts without running a language model to perform the search. Synonyms, paraphrases and questions whose useful terms never appear in the source need separate evaluation or additional retrieval logic.

The term-to-curve mapping supplies an identifier for a term. It does not establish that nearby curve points represent similar meanings, or that a high-scoring record is factually correct.

A structured query can contain several fields, but the underlying search combines matching postings and scores the candidates. Supplying two fields does not impose an SQL-style requirement that every returned document match both. Applications needing exact constraints must apply explicit filters.

The Wikimedia facade adds filters for language, minimum pageviews, maximum visibility rank and the presence of a Wikidata identifier. It can collapse records with the same non-empty Wikidata identifier to one representative, which is useful when a supplied collection contains language variants.

Two details matter when designing an interface:

• Filtering and entity deduplication happen after a bounded candidate fetch. A restrictive filter may return fewer results than requested even when additional matching records exist outside that candidate set. Reported match counts describe that fetched set, not a complete corpus count.
• The wikidata_id option currently contributes a search term but is not checked as an equality constraint by the filter function. It should not be presented as an exact entity filter without an additional check or a fix.

The dedicated Wikimedia service also holds one active snapshot at a time. Entity deduplication does not, by itself, establish a cross-project federation.

ECAI has two posting-proof formats. The older proof refers to an internal document number. The version 2 path uses a stable document identifier and a SHA-256 commitment to the canonical record, then builds a Merkle membership proof against the corresponding term's posting root.

Item	What a consumer can use it to check
Artifact file digest	Whether received bytes match the referenced file
Version 2 record commitment	Whether the record matches its committed content
Version 2 posting proof	Whether that document and record commitment belong under a specified term root
Trusted release reference	Which published root the consumer intended to accept

The Wikimedia artifact exports version 2 term headers. A proof-aware client would explicitly request proof_for_v2/3, verify the record commitment and membership path, and compare the result with a root obtained from a trusted release. The default search/3 header response uses the older format; clients must not silently mix the two schemes.

A valid membership proof establishes inclusion relative to the chosen root. It does not prove that the search returned every matching record, that its ranking is optimal, or that the underlying statement is true. A trustworthy answer also depends on the source, the collection's coverage and the reasoning applied to the retrieved text.

The snapshot service copies a completed snapshot into storage it owns, records its digest and persists the active pointer. This keeps search from depending on a temporary job directory. It loads a replacement context before discarding the previous one; loading remains a synchronous server operation.

One integrity gap remains visible in the reviewed source: restart restoration loads the saved path without comparing the file with its recorded digest. The existing-file branch of installation also trusts a digest-named file without rehashing it. A deployment relying on verified artifacts should close that gap and test corrupted-file rejection before describing every reload as cryptographically verified.

The 7 October integration review rechecked these search and snapshot modules. The exact-entity-filter and reload-integrity findings remain open in that source.

Snapshot hashes identify the exact bytes produced. The current snapshot serializer does not establish byte-identical output for every equivalent rebuild across insertion orders or runtime environments. Keep reproducible source selection, record commitments and byte-for-byte artifact identity as separate claims.

An evidence browser is a concrete starting point: show the source title, excerpt, URL and active corpus release beside each answer. Add proof verification where the release and record can be checked end to end. Such a browser could support public research collections, support documentation or comparisons between model-generated answers using the same evidence.

Evaluate it with known questions, misleading near-matches, restrictive filters and intentionally altered records. Measure relevance and missed results separately from proof acceptance. The repository includes Wikimedia search tests and snapshot service tests; no new runtime or relevance result is asserted here.

Read how the Wikipedia corpus is assembled and how an artifact can become a versioned release reference.