Skip to main content
Reference for config.yaml. Environment placeholders like ${VAR_NAME} resolve at runtime.

​
Top-level Keys

  • sources: connector list for ingestion jobs
  • embedding: embedding provider + model config
  • inference: LLM provider + model config for rephrase/chat tasks
  • vector_store: chunking and pgvector/HNSW settings

​
sources

Each item: 
- type: "s3" # or mediawiki / serpapi / etc connector type
  name: "account1"
  config:
    # connector-specific fields
Connector config fields differ by connector type. See connector pages for exact keys.

​
Common Connector Fields

The following fields are supported by all connector types:
FieldRequiredDefaultDescription
schedulesno3600Ingestion interval in seconds
request_delayno0Seconds to wait between outbound API requests. Increase to avoid rate-limiting (e.g. 0.1)

​
embedding

embedding:
  provider: local # local | openrouter | openai
  model_config: sentence-transformers/all-mpnet-base-v2
  embedding_dim: 768
  • provider: embedding backend
  • model_config: model id/name used by provider
  • embedding_dim: vector dimension returned by model
WARNING: once embedding dimension is set, it cannot be changed without rebuilding the index.

​
inference

inference:
  provider: openrouter # openrouter | openai
  model_config: openai/gpt-oss-20b:free
  • provider: inference backend
  • model_config: chat/rephrase model id
NOTE: inference backend is not used for vector search, only for the rephrase endpoint.

​
vector_store

vector_store:
  table_name: embeddings
  hybrid_search: true
  chunk_size: 512
  chunk_overlap: 50
  hnsw:
    hnsw_m: 16
    hnsw_ef_construction: 64
    hnsw_ef_search: 40
    hnsw_dist_method: vector_cosine_ops
  • table_name: embeddings table name
  • hybrid_search: vector + keyword retrieval toggle
  • chunk_size: chunk size for indexing
  • chunk_overlap: overlap between adjacent chunks
  • hnsw.hnsw_m: graph neighbors per node
  • hnsw.hnsw_ef_construction: index build quality/speed tradeoff
  • hnsw.hnsw_ef_search: search recall/speed tradeoff
  • hnsw.hnsw_dist_method: vector distance operator