AI-Powered Search Results Navigation with LLMs & JSON Schema
summary
An AI Filter Assistant for SDMX data is implemented with Python FastAPI, OpenAI GPT-4.1 (comparison vs GPT‑4 and GPT‑4o‑mini), and Apache Solr; it builds cached SDMX dictionaries, runs staged LLM calls (topic → country → dimension), and returns constrained JSON via JSON Schema plus constraint decoding to avoid hallucinations. Tested on 23 data flows / 24 manual queries and a smaller automated suite; results vary (not deterministic), average extra latency ~4–5s, and limitations include schema size, required fields, and occasional prompt failures.
AI-written summary · about
chapters (17)
- 0:00Introduction and team
- 3:59BIS data portal overview
- 7:58LLM search use case
- 9:47API structure and pipeline
- 12:57Data retrieval and caching
- 16:13Topic selection with LLM
- 18:20Country and dimension selection
- 20:49Prompting challenges and lessons
- 22:35Structured output with JSON Schema
- 25:48JSON Schema implementation
- 29:04JSON Schema limitations
- 30:08Testing framework and evaluation
- 34:12Improvement ideas and future work
- 36:20Survey and closing remarks
- 37:02Q&A on LLM processing time
- 39:10Q&A on JSON reasoning order
- 41:09Closing and applause
related talks
38mAI-Ready Data Foundation Using Amazon OpenSearch
39mExperiments with OpenSearch and AI
37mAgentic AI Meets Search: Real-World Use Cases With...
30mBeyond Keywords: Building an AI-Powered Search That Reads Your Mind
35mAugmented Retrieval Products with GAI models
42mOpenSearch With Gaia: Building Next-Gen AI-Driven Search and Analytics Workflows
41mFrom Natural Language to Structured Solr Queries using LLMs
38mA Visual AI Search Flow Builder for OpenSearch: Build GenAI and Vector...