# Selective re-embed index

`com.tnsai.intelligence.rag.vector.SelectiveReembedIndex` is a local
vector index that re-embeds only chunks whose content hash changed.
It is `@since 0.14.0` in TnsAI 0.14.0 (`TnsAI@1291fe9e`,
TAN-3660 / PR #150). It ships in Maven Central `0.14.1`.

This is the laptop / default ingest path. Qdrant and pgvector stay the
production `com.tnsai.memory.advanced.VectorIndex` backends. This type
does **not** implement `VectorIndex`, and nothing in Role RAG opens it
from an env URL or `@VectorMemory.provider`.

Storage vs a naive full reindex is the unique-hash set, not a
marketing 97%. Recompute cost on a one-file edit is that file's new
hashes, not the whole corpus.

## What it does

```java
SelectiveReembedIndex index = new SelectiveReembedIndex(embeddings);
ReindexStats first = index.syncSource("a.md", List.of("alpha", "shared"));
ReindexStats again = index.syncSource("a.md", List.of("alpha changed", "shared"));
List<ScoredChunk> hits = index.search("alpha", 3);
```

`syncSource(sourceId, chunks)` replaces one source's ordered chunk
texts. Each chunk is keyed by SHA-256 of its UTF-8 bytes:

1. a hash that already has a vector is **reused**
2. a new hash calls `EmbeddingFunction.embed` once and stores that
   vector
3. a hash dropped from this source is **removed** only when no other
   source still owns it
4. consecutive chunks on the same source become neighbors

`ReindexStats` reports `considered`, `reused`, `embedded`, and
`removed`. Duplicate payloads across sources share one stored
embedding. `size()` is the unique-hash count. `uniqueEmbeddingRatio()`
is unique / naive; `1.0` means no sharing.

`search(query, topK)` embeds the query and ranks unique stored
vectors by cosine similarity. Hits are `ScoredChunk` records
(`contentHash`, `sourceId`, `text`, `score`). Dimension is pinned on
the first embed; a later mismatched length fails loud.

## Not this page

- [Qdrant](qdrant.md) and [pgvector](pgvector.md) — production
  `VectorIndex` adapters
- [Knowledge Base](knowledge-base.md) — `InMemoryKnowledgeBase` is a
  different store
- Server FILE incremental skip — SHA-256 of whole files on the
  indexer path, not this type

## Related

- [Pipeline](pipeline.md) — FILE ingest and Server `FileIndexer`
- [Embeddings](embeddings.md) — process-wide `EmbeddingFunction`