little perch
Dear LibreChat Community,
We've encountered an issue while using the "gpt-4-1106-preview" model alongside the "Azure AI Search" plugin. Specifically, when we use this model, our input query ends up containing only newlines. Strangely, this problem doesn't occur when we use either the "gpt-4-turbo-preview" or "gpt-3.5-turbo-1106" models with the same plugin.What's interesting is that this issue seems to affect only queries written in Japanese, English queries work as expected
To provide more context, the error occurs after the following function call in the source code: PluginsClient.js line 329
Our understanding is that this problem is related to LLM chaining, specifically when using the agent and tools. Here's our reasoning:
- If we don't use the "Azure AI Search" plugin (or any tool for that matter), the program interacts directly with the model's chat completion API. This bypasses the LLM chaining mechanism.
- In this scenario, the issue doesn't occur, suggesting that the problem might arise during processing within the LLM chaining pipeline when the "Azure AI Search" plugin is involved.
We suspect that after the LLM run completes, the query keyword gets transformed into just newlines. We're wondering if this behaviour is specifically tied to LangChain.
For example, consider the following query:
東京のIT労働環境に関する質問を探す
検索結果から新しい質問と回答を生成する
＃出力
【質問】：
【答え】：
After the LLM run completes, we observe that this query becomes {""\n\n"}.
@tulip mulch can you please respond to this issue? Are there are any known issues related to this behaviour with the "gpt-4-1106-preview" model and LLM chaining in LibreChat? Any insights or suggestions to resolve this problem would be greatly appreciated.
Thank you.
tulip mulch