Peter Werry builds a retrieval systemRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task. for engineering information such as pull requests, tickets and conversations. A request for recently merged work contains author, date and state constraints that semantic similarity alone cannot reliably enforce. His workshop treats relational context as a complement to vector retrievalVector search finds items by comparing numerical representations of their features or meaning., not a reason to discard it.
Peter Werry demonstrates schema inference from sampled documents, entity resolutionEntity resolution identifies records from different sources that refer to the same real-world person, company, account, or object., a model-generated query plan and controlled database execution. Trusted code validates permitted operators, hides internal fields and applies tenant restrictions rather than leaving security boundaries to model instructions. Validation or execution errors feed bounded retries, with full-text retrieval available as a fallback.
Peter Werry discusses continuous ingestionA data pipeline moves information through ordered collection, transformation and delivery steps for use by software or AI., schema selection at larger scale and the combination of lexical and semantic search. The prototype illustrates the architecture, but sampled schemas are not completeness guarantees and production systems require stronger controls. Questions explore how these techniques can help explain engineering decisions across code reviews and chat.
Watch on YouTube
