Skip to main contentThis page defines the core concepts you’ll use across Moss.
Index
Structure that powers fast, local search. You add documents, we build an efficient index for sub-10ms queries.
Document schema
id (string), text (string), metadata (optional string map)- Upserts replace matching
ids; keep ids stable for updates
Embeddings
Semantic vector representations of text. Can be generated locally or via a remote model, then stored alongside your index.
Chunking
- Aim for ~200–500 tokens per chunk; overlap 10–20%
- Smaller chunks improve recall; overlap preserves context continuity
Retrieval
How results are fetched. Options include vector similarity, keyword, and hybrid methods with scoring and reranking.
Retrieval knobs
top_k: number of results to returnalpha: blend semantic (1.0) vs keyword (0.0); defaults semantic-heavy- Filters: constrain by metadata (e.g., category/lang)
- Rerank: reorder top-k for precision
Storage & Sync
Persist indexes locally (desktop/mobile). Optionally sync segments to the cloud for backup and sharing.
Models
moss-minilm (default): fast, lightweight for edge/offlinemoss-mediumlm: higher accuracy with reasonable performance
Authentication
Used for optional cloud features like syncing or hosted embedding models. Local mode requires no network access.
Lifecycle
- Create index → upsert docs → load → query → delete when done
- Supports multiple indexes per project
- Sub-10ms local queries (hardware-dependent)
- Sync is optional; compute stays on-device
