Bug Description

When using reasoning-capable LLMs (e.g. MiniMax-M2.5, DeepSeek-R1, QwQ) as the VLM backend,
<think>...</think> reasoning blocks in model responses are stored verbatim into file summaries
and directory overviews/abstracts. This pollutes semantic search results and wastes token budget
during L0/L1 context loading.

Root cause: openviking/models/vlm/backends/openai_vlm.py returns message.content directly
without any post-processing. Many reasoning parsers (e.g. vLLM's minimax_m2_append_think,
deepseek_r1) keep <think> blocks inside message.content.

Steps to Reproduce

1. Deploy a reasoning LLM with vLLM (e.g. MiniMax-M2.5-AWQ with
   reasoning_parser: minimax_m2_append_think)
2. Configure it as the VLM backend in ov.conf
3. Run ov add-resource <directory> to import documents
4. Inspect generated summaries/abstracts via ov search or direct storage query

Environment

• vLLM: 0.17.0
• LLM: MiniMax-M2.5
• vLLM reasoning_parser: minimax_m2_append_think
• vLLM max_model_len: 65536

Expected Behavior

Stored summaries and abstracts contain only the final output content.
<think>...</think> reasoning blocks should be stripped before storage.

Actual Behavior

Summaries and abstracts contain the full <think> block

Minimal Reproducible Example

Error Logs

OpenViking Version

0.2.6

Python Version

3.11

Operating System

macOS

Model Backend

OpenAI

Additional Context

No response