Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket
A question about payment rules can require one table row, an entire section or several rulebooks. Eleonora Drykina and Nina Žižakić describe a RAG assistant built for an anonymous Belgian payments client whose source material spans about 270 PDFs and 6,000 pages. Much of the answer bearing information sits in tables. Their indexing pipeline preserves section paths and chunk order so a larger passage can be reconstructed, and converts tables into CSV while retaining the relationship between headings and values. Combining dense retrieval with BM25 raises the reported retrieval hit rate from 92% to 97% on 63 questions; that measures retrieval, rather than complete answer correctness. The serving agent can widen its search through four tools, moving from search results to sections, full documents and multiple rulebooks. Limits prevent an oversized rulebook from flooding the context, and the prompt asks the agent to reconsider whether another tool call adds evidence. A median answer takes 38 seconds, which makes the interface part of the engineering problem. Streaming Markdown separately from a completed source card avoids waiting for one large structured object before showing anything. The speakers also explain traces, citation checks and Promptfoo evaluations, including cases where an automated judge's apparent regression did not survive inspection of the source. Their early internal rollout reached 34 users and roughly 900 questions; external release remained planned. The practical lessons concern reconstructable evidence, bounded retrieval, visible progress and evaluation that can be challenged against the documents. Speaker info: Eleonora Drykina: - https://www.linkedin.com/in/eleonora-drykina-3a1212233/ - https://www.elu.sh/ Nina Žižakić: - https://www.linkedin.com/in/ninazizakic/ Timestamps: 00:00 - A payment rulebook assistant in production 2:42 - PDFs, tables and the source corpus 5:47 - Indexing and the system architecture 7:05 - Reconstructable sections without overlapping chunks 9:52 - Turning complex tables into usable evidence 11:42 - Dense retrieval, BM25 and reciprocal rank fusion 13:26 - Four tools that widen the search 14:22 - Context limits and tool call discipline 15:32 - Latency and streaming the answer 17:38 - Traces, citations and evaluation 20:37 - Early internal use and planned external release 22:46 - Lessons from the build




Join the discussion
Sign in to join the discussion
Sign in