System Behavior & Mechanics
AI agents execute tools by generating structured JSON payloads based on injected tool schemas. Lower-parameter or highly quantized local models frequently fail to generate valid JSON schema matches, leading to tool-call failures.
Issue: LLM Hallucinated a Tool Call
The model pretends to execute a tool and outputs a text response, but no execution “thought-chain” is displayed in the UI.
Remediation
- Clear Context: Execute
/resetin the chat to clear history and prevent past outputs from biasing the JSON generator. - Upgrade Model: Deploy a larger parameter model or a model with a lower quantization level (higher precision).
Issue: LLM Refuses to Execute a Specific Tool (e.g., Web Scraping)
The model explicitly states it cannot or will not call a specific tool.
Remediation
- Reduce Prompt Context: Disable all unused tools in the workspace settings to minimize the prompt window size.
- Switch Models: Deploy a model with less restrictive alignment training.
- Reset Session: Execute
/resetand resubmit the prompt.
Issue: LLM Fails to Detect or Execute Any Tools
The model fails to discover tools or fails to output the exact JSON schema required for execution.
Remediation
- Isolate Agent Model: Configure the agent to use a high-parameter cloud-based model (or a model fine-tuned specifically for function calling/JSON generation) while retaining local open-source models for standard chat.
- Minimize Tool Overhead: Disable unused tools to reduce schema payload in the system prompt.
- Reset Context: Execute
/resetto clear corrupted chat histories affecting JSON generation.