Retrieval-Augmented Generation Over the L2 Data Fabric

Copilots ground every claim by retrieving from the L2 data fabric — shot histories, design documents, fault records — rather than relying on parametric memory.

Grounding over recall

Kronos copilots do not answer from a model's parametric memory where facts matter. They retrieve. Retrieval-augmented generation (RAG) over the L2 data fabric pulls the specific shot histories, design documents, calibration records, fault logs, and procedure templates relevant to a request, and the copilot reasons over that retrieved evidence with citations. This is what keeps advice tied to the actual machine and its canon rather than to a plausible-sounding hallucination.

What is indexed

• Shot and scenario histories (waveforms, outcomes, twin runs)
• Design documents and the frozen canon (Q_sci 3.076, 85.0 MW, 26.49 T plug, ...)
• Fault and maintenance records with root-cause attributions
• Approved procedure and runbook templates, standing orders

Retrieval spans two indices: a structured query path into the L2 shot database and feature store for exact, filterable facts, and a vector index for semantic recall over documents. A copilot typically issues both — structured retrieval for numbers, vector retrieval for the surrounding engineering rationale.

retrieve(query, machine):
  facts = L2.sql(structured_filter(query, machine))   # exact
  docs  = vector_db.search(embed(query), k, filter=machine)
  ctx   = rerank(facts + docs, query)[:budget]
  return ctx with provenance ids for citation

Provenance is mandatory

Every retrieved item carries a provenance identifier through the whole loop, so the final answer can cite the exact source. An unsourced assertion is treated as a defect by the grounding evaluation. Retrieval also respects data lineage and access scope from L2 — a copilot cannot retrieve what it is not authorized to see, and confidential material is firewalled from public-facing outputs.

Because both machines are design-and-simulation studies, much of what is retrieved today is twin runs and design documents rather than real-shot data. As the breeder is commissioned toward FOAK, real shot histories enter the same fabric and the same retrieval path, with no change to the copilots' grounding contract. See chunking and embedding and citation grounding.