Resolved

vLLM model fails to parse European dates and tool calls in agentic workflows

Python
AI/ML
vLLM
LLM
NLP
Debugging
Date Parsing
Tool Calling
Model Configuration
Quantization
Regex
Backend

DateTime parser silently failed on DD.MM.YYYY format while vLLM 0.[IP] lacked proper tool call and reasoning configuration | Caused hallucinated responses in horse booking system with Qwen3 models

Issue

The system encountered failures when processing European date formats (e.g., DD.MM.YYYY) and generated hallucinated responses when using vLLM with Qwen3 models. These failures included silent parsing errors for dates and exposure of internal reasoning blocks in user-facing responses.

Culprit

DateTime Parser Failure

The regex used for parsing datetime values only supported MM/DD/YYYY format, resulting in silent failures for European date formats. This caused downstream tools to receive malformed inputs, leading to incorrect processing.

vLLM Configuration Mismatches

Incorrect parser configurations and missing model-specific settings in vLLM led to ignored tool calls and the leakage of internal reasoning blocks in responses. Additionally, quantization conflicts and undocumented arguments caused further complications.

Model Capability Gaps

Smaller models, such as Qwen3-4B and Llama-3.1-8B, exhibited hallucinations and failed to handle tool calls properly. These issues resulted in invented booking details and skipping failed tool calls without fallback mechanisms.

Solution

DateTime Parser Fix

The regex was updated to support both DD.MM.YYYY and MM/DD/YYYY formats, including logic to automatically detect and handle European date formats.

vLLM Configuration Resolution

The vLLM configuration was updated with model-specific parsers, enabling auto tool choice and reasoning capabilities. This resolved issues with ignored tool calls and reasoning leaks.

Model Selection

An evaluation of multiple models was conducted, and the Qwen3-8B-FP8 model was selected for its stable operation and ability to handle tool calls and reasoning correctly.

Critical Discoveries

  • Parser Validation: Confirmed the valid parsers for vLLM and adjusted configurations accordingly.
  • Memory Constraints: Reduced model length settings to fit available VRAM, ensuring stable performance.
  • Version Limitations: Documented undocumented features and limitations in vLLM, such as the non-existent --enable-reasoning flag.

These changes successfully addressed parsing errors, resolved model hallucinations, and ensured consistent and reliable operation of the system.