Glossary
Glossary
The vocabulary around local embeddings and caching, defined plainly.
- Embedding
- A fixed-length vector of floats that represents a piece of text in a way that puts semantically similar text close together. embedcache generates these locally.
- FastEmbed
- The Rust crate embedcache builds on to run ONNX embedding models on CPU without a hosted API. Source of the 22+ bundled models.
- ONNX runtime
- The engine that executes the exported model weights. It is the CPU-first path embedcache uses for inference — no GPU required.
- Content-hash cache
- A cache keyed by a hash of the input content (plus the model id) rather than by an arbitrary key. Identical inputs deduplicate automatically.
- Cache hit / miss
- A hit means the requested vector already exists in SQLite and is returned without inference. A miss triggers a fresh embed, which is then written to the cache.
- Model identifier
- The name of the embedding model (e.g. BGESmallENV15). It is part of the cache key, because the same text under a different model is a different vector.
- Chunking
- Splitting a document into smaller pieces before embedding. embedcache offers a whitespace words chunker by default and optional LLM-based semantic chunkers.
- Vector database
- A system that indexes and searches embeddings (Qdrant, pgvector, LanceDB, Milvus). embedcache produces vectors for these; it is not one itself.
- Dimensions
- The length of an embedding vector. Bundled models range from 384 (MiniLM, BGE small) to 1024 (BGE large). Higher can capture more nuance at higher cost.
- Backfill
- Embedding an existing corpus in bulk. Against a hosted API this is rate-limited; locally it is bounded only by your CPU, and cache hits skip work entirely.