Retrieval config

Every knob in workspace/<org>/retrieval.yaml. The full pipeline is pluggable — swap embedders, rerankers, chunking strategies without touching code.

Status: First-party pgvector + Postgres FTS hybrid pipeline is live. Configuration via workspace/<org>/retrieval.yaml.

Full example

embedder:
  provider: openai
  model: text-embedding-3-large
  dimensions: 3072
  batch_size: 96

chunking:
  strategy: recursive # recursive | semantic | fixed
  size: 1200
  overlap: 200
  respect_boundaries: [heading, paragraph]

vector:
  store: pgvector
  index: hnsw # hnsw | ivfflat
  m: 16
  ef_construction: 64
  ef_search: 80

keyword:
  store: postgres_fts
  language: english
  boost_fields: {title: 2.0, body: 1.0}

hybrid:
  method: rrf # reciprocal rank fusion
  k_constant: 60
  vector_weight: 1.0
  keyword_weight: 0.6

rerank:
  enabled: true
  provider: voyage
  model: rerank-2
  top_n_input: 40
  top_n_output: 8

query:
  k_nearest: 8
  similarity_threshold: 0.3
  max_context_tokens: 8000

# Per-source / per-domain overrides
overrides:
  sources:
    zoom:
      chunking: {size: 2000, overlap: 300} # longer transcripts
      rerank: {top_n_output: 4}
  domains:
    sales:
      embedder: {model: text-embedding-3-small} # cost optimization

Providers

Auditability

Every skill_run persists:

• retrieval_config_sha — the hash of the YAML that was active
• retrieval_hits — the full set of (document_id, score, rank) tuples returned

Makes "why did the agent retrieve X on March 3rd" a SQL query, not a debugging session.