#10550: feat(memory-lancedb): local embeddings via node-llama-cpp

## feat(memory-lancedb): local embeddings via node-llama-cpp > **AI-assisted:** This PR was developed with Claude Code/Opus 4.6. I understand what the code does, have reviewed all changes, and tested on Ubuntu Linux. The full gate (`pnpm build && pnpm check && pnpm test`) passes. ### Summary Adds offline embedding support to the `memory-lancedb` plugin using [node-llama-cpp](https://node-llama-cpp.withcat.ai/), alongside dimension mismatch detection and a reindex command for safe provider switching. The plugin previously required an OpenAI API key for all embedding operations. This PR adds a `local` provider that runs GGUF embedding models entirely on-device — no API key, no network, no ongoing cost. Users who hit rate limits, work in air-gapped environments, or simply prefer keeping data local can now use long-term memory without any cloud dependency. ### What changed **Local embedding provider** - `LocalEmbeddingProvider` class wraps node-llama-cpp's `EmbeddingContext` - Supports GGUF model files (local path or `hf:` URL for auto-download) - L2-normalized output vectors for consistent similarity scoring - Lazy model loading — downloaded and initialized on first use **Provider auto-detection** - Models ending in `.gguf` or starting with `hf:` auto-select `local` - Models starting with `text-embedding-` auto-select `openai` - Explicit `provider` field takes precedence **Dimension mismatch safety** - On startup, reads the Arrow schema from the existing LanceDB table and compares vector dimensions against the current config - Throws `DimensionMismatchError` with a clear message if they differ (e.g. switching from OpenAI 1536-dim to local 768-dim) - Instance stays un-initialized on mismatch — no silent corruption **`openclaw ltm reindex` CLI command** - Opens the old table without dimension validation, reads all entries - Drops and recreates with the new dimension, re-embeds each entry through the current provider - Reports progress and success/failure counts **Documentation** - New plugin doc: `docs/plugins/memory-lancedb.md` — config, local setup, CLI reference, provider switching - Added to sidebar and linked from plugin list ### Config examples ```json // OpenAI (existing, unchanged) { "embedding": { "apiKey": "${OPENAI_API_KEY}", "model": "text-embedding-3-small" } } // Local (new) { "embedding": { "provider": "local", "model": "hf:nomic-ai/nomic-embed-text-v1.5-GGUF/nomic-embed-text-v1.5.f16.gguf" } } ``` ### Files changed | File | Change | |------|--------| | `extensions/memory-lancedb/index.ts` | LocalEmbeddingProvider, DimensionMismatchError, reindex CLI, schema validation | | `extensions/memory-lancedb/config.ts` | Provider detection, local config parsing, dynamic dimensions | | `extensions/memory-lancedb/index.test.ts` | 15 new tests (provider detection, dimensions, config, mismatch) | | `extensions/memory-lancedb/openclaw.plugin.json` | Extended schema and UI hints for local provider | | `extensions/memory-lancedb/package.json` | Added node-llama-cpp as optional dependency | | `docs/plugins/memory-lancedb.md` | New plugin documentation | | `docs/docs.json` | Sidebar entry | | `docs/plugin.md` | Link to plugin doc | ### Testing **Tested on:** Ubuntu Linux **Degree of testing:** Lightly tested locally, fully tested via automated suite. - [x] `pnpm build && pnpm check && pnpm test` all pass (5452 tests, 0 failures) - [x] Provider auto-detection: `.gguf` → local, `text-embedding-*` → openai, explicit override - [x] Vector dimension calculation: 1536 for OpenAI small, 3072 for large, 768 default for local - [x] Config rejects missing apiKey for OpenAI, accepts omitted apiKey for local - [x] Dimension mismatch throws on init when existing DB has different dimensions - [x] `initializeUnchecked()` bypasses validation for reindex access - [x] `recreateTable()` drops old table and creates with new dimensions <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> - Adds a new `local` embedding provider backed by `node-llama-cpp`, plus provider auto-detection and updated config parsing/UI hints. - Introduces LanceDB vector dimension validation on startup, raising a dedicated `DimensionMismatchError` when an existing table doesn’t match the configured embedding dimension. - Adds `openclaw ltm reindex` CLI command to drop/recreate the table and re-embed all stored memories when switching providers. - Updates docs and plugin schema to document/configure the new provider options and workflows. <h3>Confidence Score: 3/5</h3> - This PR is close to mergeable, but has a couple of config/CLI logic issues that can cause incorrect reindexing or guaranteed dimension mismatches in valid configurations. - Core changes are straightforward and covered by new tests, but the runtime code currently computes vector dimensions from `embedding.model` even when local embeddings use `local.modelPath`, and the reindex command can reuse a stale embedding provider instance. Both can lead to incorrect data in the DB or startup failures after provider switching. - extensions/memory-lancedb/index.ts (provider selection/vectorDim computation and ltm reindex embedding provider lifecycle) <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->