Weaviate paper: LLMs misuse LIKE instead of search operators when querying databases
CShorten30 · x · 2026-09-14
A Weaviate paper, Querying Databases with Function Calling, proposes a unified tool definition for database querying supporting search, filters, aggregation and multi-collection routing. Using a Gorilla-adapted synthetic pipeline (DBGorilla benchmark, compared to BIRD/Spider/WikiSQL) across 8 LLMs from 5 families, they show Claude 3.5 Sonnet and GPT-4o handle it well — but LLMs frequently use LIKE or exact text match operators for natural language commands requiring a SEARCH operator.
Related event: Paper: LLMs Prefer LIKE Over Semantic Search When Querying Databases(2 posts)→
More from coding & agent
- InnerTone: an OpenAI hackathon demo builds a browser music "compiler" with Codex — cedric_chee · 2026-09-14
- Claude Prompting Levels 6-7: Iteration Feedback and Autonomous Workflows with Examples — ZabihullahAtal · 2026-09-14
- A 7-Level Prompting Framework for Getting Better Results from Claude — ZabihullahAtal · 2026-09-14
- Qwen Code v0.23.4 ships Goal budgets, cross-session messaging endpoint, and MCP fixes — qwen-code-ci-bot · 2026-09-14
- Amazon paper: LLM judges rate 57.5% of failed agent tasks as satisfactory, flip close rankings 31% of the time — dair_ai · 2026-09-14
- Building a Claude Code Template Library for Expert-Level Direct-Response Pages — boringmarketer · 2026-09-14