SemanticIndex.h header

#include <ew/app/SemanticIndex.h>

Namespace ew::app

Entry struct

struct ew::app::Entry

One indexed content object and its embedding.

Members

ew::core::ContentId ew::app::SemanticIndex::Entry::id

The indexed content object.

std::vector<float> ew::app::SemanticIndex::Entry::vector

Its stored embedding vector.

SemanticIndex class

class ew::app::SemanticIndex

An in-memory vector index over content embeddings for semantic (meaning-based) retrieval: add an embedding per content object, then search returns the closest ids by cosine similarity. This is the core of canon-aware AI retrieval (RAG) – ground an AI request on the handful of Codex entities most relevant to a query even when the whole codex would overflow the model's context. Provider-agnostic (fed by whatever embeddings endpoint the user configures); a linear scan, which is ample for a project's entities; not thread-safe.

Members

void ew::app::SemanticIndex::add(ew::core::ContentId id, std::vector< float > vector)

Adds vector for id, replacing any existing entry for that id. A null id or an empty vector is ignored.

void ew::app::SemanticIndex::remove(ew::core::ContentId id)

Removes the entry for id if present.

void ew::app::SemanticIndex::clear()

Drops all entries.

std::size_t ew::app::SemanticIndex::size() const

The number of indexed entries.

bool ew::app::SemanticIndex::contains(ew::core::ContentId id) const

True when id has an entry.

std::vector< ew::core::ContentId > ew::app::SemanticIndex::ids() const

The ids currently indexed, in insertion order (useful for spotting what still needs embedding, or what has gone stale).

std::vector< float > ew::app::SemanticIndex::vectorFor(ew::core::ContentId id) const

The stored embedding for id, or an empty vector when id is not indexed. Lets a caller find items similar to an already-indexed one without re-embedding it (feed the result to search).

std::vector< SemanticMatch > ew::app::SemanticIndex::search(const std::vector< float > &query, int topK) const

The topK entries most similar to query by cosine similarity, most-similar first. Zero-similarity (unrelated) entries are skipped, so an off-topic query returns fewer – or no – results. Returns empty when query is empty or topK is not positive.

SemanticMatch struct

struct ew::app::SemanticMatch

One retrieval result: the matched content id and its cosine similarity to the query (higher is closer). Results are returned ranked most-similar first.

Members

ew::core::ContentId ew::app::SemanticMatch::id

The matched content object.

float ew::app::SemanticMatch::score = 0.0F

Cosine similarity of its embedding to the query, in (0, 1].

bool operator==(const SemanticMatch &, const SemanticMatch &)=default

Matches compare equal when both fields match.