Hybrid retrieval

Lamina uses retrieval to choose relevant Workflow roots and locate source symbols. Retrieval never defines product behavior. After selection, graphd traverses the canonical product graph for ordered Operations, Actors, authority, states, Invariants, Scenarios, Surfaces, risks, and proof requirements.

Two stores, two authority levels

graphd is the only process that writes either Ladybug database. The native worker reads repository files, creates syntax-aware chunks and embeddings, and sends authenticated batches to graphd.

Automatic synchronization

lamina work prepare compares the active index manifest with:

• GraphVersion;
• source and repository revision;
• branch and worktree identity;
• retrieval schema version; and
• model digest.

If any value changed, the worker creates a pending generation. Unchanged documents reuse their content-addressed versions; changed documents are upserted; removed files or symbols are absent from the new membership. graphd activates the generation only when committed membership and digest match the declared expected values. An interrupted generation cannot replace the last complete current generation.

Ranking

Workflow selection gives an exact id or alias match absolute priority. Otherwise it combines Ladybug BM25 and cosine rankings with fixed reciprocal-rank fusion, filters inactive roots, and applies small deterministic boosts for direct Actor, Persona, Entity, Surface, state, or Operation terms. The result is explicit:

• selected;
• multi_workflow;
• ambiguous; or
• new_workflow_required.

Source localization uses the same BM25, dense, and fusion policy over syntax-aware chunks. Packet v5 records file, enclosing symbol, line range, score, and short match reasons. Overlapping chunks from one symbol are deduplicated.

An explicit --workflow <id> bypasses semantic Workflow selection, but it still performs exact graph traversal and source retrieval.

Offline runtime and integrity

The release pins jinaai/jina-embeddings-v2-base-code to one immutable revision and publishes its dynamic-INT8 ONNX weights as a separate checksummed asset. The platform worker contains ONNX Runtime, the tokenizer, and matching Ladybug FTS and vector extensions. Installation extracts these assets into the same private versioned runtime as the worker.

Lamina never downloads a model or extension during normal execution. Missing, corrupt, or mismatched assets return LAMINA_RETRIEVAL_INTEGRITY with reinstall guidance.

Diagnostics and recovery

lamina context status
lamina context rebuild

status reports generation, freshness, model and index digests, Workflow and source-chunk counts, pending state, and the last failure. rebuild invalidates the current derived generation for this branch/worktree identity and reconstructs it synchronously.

Automatic Workflow selection fails closed when dense retrieval is unavailable. With an explicit Workflow id, exact graph closure may continue and source localization can use BM25 alone; packet v5 marks that state lexical_degraded.

Source comments and source text remain untrusted retrieval data. They cannot change the selected Workflow’s graph instructions or packet obligations.