Enterprise AI Bootcamp Demo 5

Engineering Research Assistant

From a published finding to an executed experiment with a measured result, in one sequence. The corpus is 0 real arXiv papers, every identifier resolved against the arXiv API, plus a small set of clearly-labelled synthetic internal engineering notes.

Before any of this is a retrieval problem it is a licensing problem. Three of these papers cannot legally enter the index at all, and forty can be cited but not served. A corpus build that ignores that ships a liability, not a feature.
Index built 2026-08-10T07:23:47. 1822 passages from 90 arXiv papers and 12 synthetic internal notes. Embeddings: Snowflake/snowflake-arctic-embed-s (ONNX int8, Apache-2.0), loaded from Demo 1. One embedding stack across both demonstrations. Demo 1 harness live model configured
Retrieval, extraction, every statistical test and the backlog are model-free. Only the synthesis paragraph on screen 2 uses a model, and it degrades to extraction.

Licence resolution

There is no such thing as "the arXiv licence". Each submission carries its own, and the Atom API does not return it — this build resolves every paper individually through the OAI-PMH arXivRaw interface, at one request every 3.2 seconds with a descriptive User-Agent, as arXiv's terms of use require. Four different licences appeared across 0 papers.

LicencePapers
Identifiers requested
0
Metadata resolved
0
PDFs downloaded
0
Failures
0
Rate limit honoured
0 s between requests

A PDF is downloaded only when the licence allows the full text into the index. For a link-only paper there is nothing to gain from holding it, and for a no-derivatives paper chunking it would itself be the infringement.

What is in the index, and what is not

Internal engineering notes synthetic

Retrieval has to span internal project reports as well as external papers, because that is the actual enterprise question: does what we measured agree with what was published? These 12 notes are invented for this demonstration. They are marked synthetic in the record, inside the indexed text itself so the marker survives into any prompt, and on every citation in the UI.

IDTitleTeamDate Type
ENG-2411-017Cross-encoder reranking on the maintenance-manual index: latency measurementsPlatform Retrieval2024-11-08benchmark
ENG-2502-004Reranking did not help on our short factoid queriesPlatform Retrieval2025-02-19experiment-report
ENG-2503-011Semantic chunking trial: cost was real, the gain was notKnowledge Systems2025-03-27experiment-report
ENG-2505-023Incident: the assistant quoted a torque value from the wrong compressor variantField Systems2025-05-14incident
ENG-2506-002Why we rejected GraphRAG for the maintenance corpusKnowledge Systems2025-06-03design-note
ENG-2507-009Judge agreement audit on the answer-quality metricEvaluation2025-07-22experiment-report
ENG-2509-014Structure-aware extraction: what it actually bought usKnowledge Systems2025-09-09experiment-report
ENG-2510-006Embedding model selection under the CPU budgetPlatform Retrieval2025-10-15benchmark
ENG-2511-021Postmortem: our recall number was measured against the wrong gold setEvaluation2025-11-26postmortem
ENG-2601-008Standing constraint set for retrieval work, 2026Architecture2026-01-12design-note
ENG-2602-013Long-context experiment: cost per answer, not accuracy, was the blockerArchitecture2026-02-24experiment-report
ENG-2604-002Retrieval evaluation set: composition and known gapsEvaluation2026-04-07design-note