Embeddings and vector stores¶
The embedding model and the LanceDB or Qdrant backend behind retrieval.
Embeddings¶
EMBEDDING_PROVIDER¶
Where vector-store embeddings are computed: huggingface runs a local sentence-transformers model, openai and ollama call a service.
EMBEDDING_MODEL_NAME¶
Embedding model identifier used by the selected provider. Spelled with the org prefix so it shares one SharedEncoder slot with AGG_EMBEDDING_MODEL and CHUNK_EMBEDDING_MODEL when aligned.
EMBEDDING_API_KEY¶
Provider API key for hosted embedding services.
EMBEDDING_BASE_URL¶
Provider base URL (for Ollama-compatible endpoints).
EMBEDDING_DIMENSION¶
Expected dense embedding vector size for core and neighborhood vectors.
EMBEDDING_BM25_MODEL_NAME¶
fastembed SparseTextEmbedding model id for the BM25 sparse lane.
EMBEDDING_QUERY_PREFIX¶
Prefix prepended to text embedded as a query. Asymmetric retrieval models underperform their spec without it — BGE wants 'Represent this sentence for searching relevant passages: ', E5 wants 'query: '. Empty (default) suits the symmetric paraphrase model. Part of the stored embedding contract: changing it requires a reindex.
EMBEDDING_DOCUMENT_PREFIX¶
Prefix prepended to text embedded as a document during indexing (E5 wants 'passage: '; BGE wants nothing). Part of the stored embedding contract: changing it requires a reindex.
LanceDB¶
LANCEDB_ENABLED¶
Enable embedded LanceDB when QDRANT_URI is unset. Uses a local directory via lancedb.connect(data_dir).
LANCEDB_DATA_DIR¶
Local filesystem directory passed to lancedb.connect(...) (supports ~ expansion).
LANCEDB_ONTOLOGY_TABLE¶
Lance table for ontology atom vectors; derived like FUSEKI_DATASET, and replaced the same way by ontocast serve and ontocast process.
LANCEDB_FACTS_TABLE¶
Lance table reserved for future fact vectors; created on init.
Qdrant¶
QDRANT_URI¶
Qdrant server URL, such as http://localhost:6333. Setting it selects Qdrant as the vector store.
QDRANT_API_KEY¶
API key for a Qdrant server that requires one.
QDRANT_ONTOLOGY_COLLECTION¶
Qdrant collection for ontology atom vectors; derived like FUSEKI_DATASET, and replaced the same way by ontocast serve and ontocast process.
QDRANT_FACTS_COLLECTION¶
Qdrant collection reserved for future fact vectors; created on init.
QDRANT_GRPC_PORT¶
Qdrant gRPC port, used when QDRANT_USE_GRPC is on.
QDRANT_USE_GRPC¶
Talk to Qdrant over gRPC instead of HTTP.
QDRANT_VECTOR_SIZE¶
Vector size override. When set, must equal EmbeddingConfig.dimension; when unset, the embedding dimension is used.
QDRANT_DISTANCE¶
Qdrant vector distance when creating collections (Cosine, Dot, Euclid, Manhattan; same as qdrant_client Distance).
QDRANT_UPSERT_BATCH_SIZE¶
Batch size used for Qdrant upsert operations.
QDRANT_TIMEOUT_SECONDS¶
Per-request timeout for Qdrant calls, in whole seconds (the client accepts nothing finer). Without one, an unreachable or hung Qdrant blocks a pipeline worker indefinitely.