How to force an LLM to use tools? DeepSeek V4 ignores search instructions
ihatebeinganonymous · reddit · 2026-08-24
The author developed a Telegram bot using DeepSeek V4 Flash but found the model often ignores system prompts to use tools for factual questions (e.g., France's World Cup qualification history), relying on internal weights instead. They seek a more mechanical way to enforce tool usage before answering to ensure verifiability. An anecdotal note suggests the new "Flash Vision Exp" variant may be more compliant with tool calling.
More from coding & agent
- Agent Memory Should Track Outcomes, Not Just Steps: Adding Success Counters to Workflows — No_Advertising2536 · 2026-08-24
- AI Agents struggle to fail fast, often avoiding error responsibility — mattpocockuk · 2026-08-24
- MCP servers criticized as new grift wrapping scripts in paywalled APIs — jasonkneen · 2026-08-24
- Dev builds ContextDB, a local memory database to prevent coding agents from forgetting project decisions — Pretty-S · 2026-08-24
- Open-Source 14-Agent System Runs 24/7 Company on Laptop, No Vector DB — thisdudelikesAI · 2026-08-24
- Agno Builder Now Supports Visual Construction of Agents and Code Export — pritisinghhhh · 2026-08-24