Hacktoberfest 2026：維護者為十月標記出來的 issue，仍然開放、適合新手。 瀏覽 Hacktoberfest issue

[Bug]: OpenAI-compatible tool calls are returned as plain text when routed through PAIR

未關閉
  • 哪裡不對？
#94 7 則留言 0 個 reaction 已指派 1 人 在 GitHub 檢視

維護者通常 5 天內回覆

@kjlubick 已經在處理了。

開始於 2026年10月1日。

評估

難度
4/5
預估耗時
3-5 天
新手友好度
55/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
go, ollama
領域
api, backend

研究方向

從 Ollama 代理的 /v1/chat/completions 進入點開始，比較存在工具定義時請求和回應的處理方式。使用文件中的 curl 請求分別針對 PAIR 和直接針對 Ollama 進行重現，然後驗證 PAIR 保留了 message.tool_calls 和 finish_reason: tool_calls，而不是將呼叫作為文字返回。

由索引模型根據 Issue 內容生成。

描述

bug
PAIR version or commit

0.1.1-463

Affected component

Ollama proxy

Environment

OS: Ubuntu 26.04.1 LTS on model host; Linux on NVIDIA DGX Spark PAIR client
Architecture: x86_64 model host; arm64 PAIR client
GPU and driver: AMD Radeon 8060S / Ryzen AI Max+ 395 on model host; NVIDIA GB10 on PAIR client
Engine and version: Ollama 0.33.2
Model: qwen3.6:27b
Cluster size: 4 nodes

Steps to reproduce

  1. Configure a PAIR cluster with an Ollama node serving qwen3.6:27b.

  2. Send an OpenAI-compatible POST request to:
    /v1/chat/completions

    The request contains multiple OpenAI function/tool definitions and asks the model to work on a task for which it should call the kanban_show function.

  3. Send the request through the PAIR Ollama proxy.

  4. Send the exact same JSON request body directly to the Ollama server hosting qwen3.6:27b, bypassing PAIR.

  5. Compare finish_reason, message.content, and message.tool_calls in the two responses.

Repeated control test:

  • Direct to Ollama: 3/3 requests returned a native tool_calls object.
  • Through PAIR: 3/3 requests returned finish_reason: stop with no tool_calls object. The intended function call instead appeared as ordinary assistant content.

The behavior is reproducible with curl and does not require an agent framework.

Expected behavior

When an OpenAI-compatible chat completion request containing tool definitions is routed through PAIR, native tool-calling behavior should be preserved.

For this request, the response should contain message.tool_calls with a call to kanban_show, and finish_reason should be tool_calls, matching the response obtained when the identical request is sent directly to the Ollama server.

Actual behavior

When the identical request is sent through the PAIR Ollama proxy, the native OpenAI-compatible tool call is not returned.

In repeated testing, PAIR returned:

  • finish_reason: stop
  • message.tool_calls: null / absent
  • The intended kanban_show call embedded in message.content as ordinary text.

Three consecutive PAIR requests produced textual representations such as:

  1. Python-like syntax:
    kanban_show(task_id="t_65ba4229")

  2. Bracket syntax:
    [kanban_show][0]

  3. XML-like syntax:
    <invoke>kanban_show(task_id="t_65ba4229")</invoke>

By contrast, three consecutive requests using the identical JSON request body sent directly to Ollama returned native OpenAI-compatible tool calls:

finish_reason: tool_calls

with message.tool_calls containing a function call to kanban_show.

The same JSON payload succeeds when sent directly to the Ollama server on the model host, so the failure is introduced only when the request is routed through PAIR.

This also prevents OpenAI-compatible agent/tool frameworks from recognizing and executing the requested function when the request is routed through PAIR.

Sanitized logs or screenshots
Confirmations
  • I searched existing issues for duplicates.
  • This is not a security vulnerability.
  • I agree to follow the Code of Conduct.
主要語言
Go
星號
1.6k
分支
266
平均合併
3 天 6 小時
30 天內合併 PR
22

環境準備

從這裡開始

  1. 先讀完整個 Issue，再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫，在一個分支上完成修改。
  4. 送出 Pull Request，並在描述裡引用這個 Issue 編號。

NVIDIA/Personal-AI-Router 的其他 Issue

查看 NVIDIA/Personal-AI-Router 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。