> ## Documentation Index
> Fetch the complete documentation index at: https://unch.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Index storage

> What unch writes locally and how .semsearch state is structured.

By default, local state lives under:

```text theme={null}
<repo>/.semsearch
```

You can override that with `--state-dir`.

## Files in `.semsearch`

### `index.db`

The local search index. This is a rebuildable cache artifact, not durable user data.

It stores immutable snapshots for each provider/model pair. Embedding vectors are kept in dimension-specific vector tables, so a 768-dimension local model and a 1536-dimension OpenRouter model can coexist without replacing each other.

### `manifest.json`

Tracks local state metadata, schema versioning, and optional remote binding.

### `filehashes.db`

Stores file-hash state used to skip unchanged files on warm reindex runs.

### `file_weights.json`

Local path weighting configuration used during ranking.

### `yzma/`

Local runtime libraries used for embedding inference.

## State-dir vs db-path

The CLI now treats the state directory as the primary unit:

* `--state-dir` is the preferred flag
* `--db` is a deprecated compatibility alias

That matches the real storage model better, because `index.db` is only one piece of the full local state.

## Provider and model snapshots

Each provider/model pair has its own active snapshot.

Examples:

* rebuilding `llama.cpp/embeddinggemma` does not replace `llama.cpp/qwen3`
* rebuilding `openrouter/openai/text-embedding-3-small` does not replace `llama.cpp/embeddinggemma`
* switching provider or model requires indexing once with that same provider/model before searching

`index.db` schema changes may reset old cache data automatically. If local search fails after an upgrade, rebuild with `unch index --root .`.

## Remote-published state

Remote CI publishes:

* `index.db`
* `manifest.json`
* `filehashes.db`

That allows published state to restore both the active search snapshot and the file-hash cache used by warm reindex scenarios.