ETD

• ai
• ams
• rag

AMS operators support thousands of custom SQL procedures, each handwritten and linked to others in ways that only emerge when something breaks. A support question lands. The clock starts. The operator pieces together fragments from manuals, spread-sheets, query files, and the colleague who wrote it two years ago and does not quite remember. The work is not technically impossible. It is fragmented across formats, tools, and people, and the cost of reassembling the picture lands on the same people who keep the infrastructure running.
AIxAMS is a heterogeneous Retrieval-Augmented Generation (RAG) system designed for this setting. AMS (Assistance Maintenance Support) is the post-production phase in which operators maintain enterprise IT solutions. The project addresses a practical problem: when knowledge is spread across interdependent documents, support work begins with reconstruction, not with answers. Standard RAG flattens this material into plain text and loses part of the structure that makes it usable. In this corpus, that limitation is most damaging for the SQL procedure library, its largest and most challenging component: thousands of handwritten stored procedures whose token syntax is semantically misaligned with the natural language queries operators actually pose. AIxAMS addresses this by decomposing each SQL file into intent-driven logical blocks, embedding each block from a natural-language description rather than from raw SQL tokens, and preserving the full procedure body for generation-time evidence. Tables and spreadsheets are handled as atomic nodes with cell-level annotations, and all document types are routed through a topology-aware ingestion pipeline that preserves the structure each format requires.
The system was tested on a real AMS knowledge base made of several hundred heterogeneous documents and exercised on domain-specific queries from the operational environment. Its goal is to give operators a retrieval layer that reflects how AMS knowledge is actually organized, so the system helps them recover the context needed to work without turning every question into a scavenger hunt.