Transformers.js Local Embedding Subsystem

Transformers.js serves as the default local execution runtime for generating semantic vector embeddings. It runs the quantized version of the all-MiniLM-L6-v2 model directly on the client CPU.

---

Model Pinned Integrity Hashes

For maximum security, sovseal verifies the model files on startup against pinned SHA-256 hashes to prevent local tampering or corrupted files:

---

Performance & LRU Caching

Because generating vector embeddings via ONNX model forward-passes requires CPU cycles (taking ~5–8 ms on modern hardware), sovseal implements an embedding LRU cache to bypass the model execution entirely for redundant queries.

Cache Characteristics

• Scope: Applied only to query/search operations (recall_memory). Storage operations (store_memory) always bypass the cache to ensure unique fact vectors are written to LanceDB.
• Default Capacity: 256 queries.
• Memory Overhead: Extremely low (sub-kilobyte footprint) as it only holds 384-dimensional Float32 vectors.
• Configuration: You can adjust the cache capacity by setting the SOVSEAL_EMBEDDING_CACHE_SIZE environment variable (set to 0 to disable the cache entirely).

---

CPU Resource Costs

For typical developer machines and agent host environments: