Deep mode search result in Spanish

[fniu]

[role=assistant] <DESCRIPTION>SysML v2 ofrece ventajas clave para MBSE: semántica formal, notación textual y gráfica unificadas, API REST estándar y modelado explícito de variabilidad. También mejora la claridad del modelado con la separación definición/uso, mayor soporte para requisitos, comportamiento y estructura, y mejor integración con flujos digitales y control de versiones.</DESCRIPTION>
<NAME>Ventajas de SysML v2</NAME>
<CONTENT>SysML v2 aporta una modernización profunda frente a SysML v1 y está diseñado para mejorar la precisión, interoperabilidad y automatización en ingeniería de sistemas basada en modelos (MBSE).

All settings are default (English OS: WSL). Is it a bug or a setting that is unknown?

[fniu]

Everyhing is in English in the fast mode.

[wangxingjun778]

Currently, Sirchmunk primarily supports Chinese and English scenarios. However, our internal testing confirms that Spanish is also functionally supported. Below are quick examples using FAST and DEEP modes:

If you have requirements for Spanish language support, please briefly describe your use case (e.g., types of documents to be retrieved, expected scale/volume, etc.). We will then proceed with dedicated adaptation and optimization tailored to your needs. 💯

⚠️ Note: The logs shown above are not bugs. They reflect intentional optimizations and adaptations specifically designed for Chinese and English scenarios.

[fniu]

I don't have requirements for Spanish and didn't ask for Spanish. My query is in English so I would expect the result is in the same language.

[fniu]

I re-run the same query, it now speaks Geman:

$ sirchmunk search "sysml v2 advantanges" <dir> --mode DEEP
...
2026-04-16 10:58:22,067 - httpx - INFO - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
[role=assistant] <DESCRIPTION>SysML v2 bietet gegenüber SysML v1 mehr Präzision, bessere Tool-Interoperabilität und eine klarere Modellierungsstruktur. Die Quellen heben insbesondere die formale Semantik, die kombinierte textuelle und grafische Notation sowie die standardisierte REST-API als zentrale Vorteile hervor. Zusätzlich verbessert SysML v2 Variabilitätsmodellierung, Wiederverwendung und die Integration in digitale Engineering-Prozesse.</DESCRIPTION>

[wangxingjun778]

Oh, I see! So to clarify, your issue is: the query is in English, and you expect the results to also be in English, but Spanish or other languages appear in the intermediate logging/output?

Also, could you please confirm: what LLM configuration are you currently using? (e.g., model name/version, inference mode, FAST/DEEP mode, etc.)

[fniu]

Oh, I see! So to clarify, your issue is: the query is in English, and you expect the results to also be in English, but Spanish or other languages appear in the intermediate logging/output?

Also, could you please confirm: what LLM configuration are you currently using? (e.g., model name/version, inference mode, FAST/DEEP mode, etc.)

I am using the model below and in the deep mode (see the command in my previous post).
LLM_MODEL_NAME=gpt-5.4-mini

Yes the issue is that it seems to speak arbitrary languages while my question has always been in English. Not only the intermdieate output, but also in the final result:

============================================================
Search Results
============================================================

# **SysML v2 – Vorteile gegenüber SysML v1**

## **Kurzüberblick**
SysML v2 bietet im Vergleich zu SysML v1 vor allem **mehr formale Präzision**, **bessere Tool-Integration** und eine **klarere Modellstruktur**. Besonders relevant sind die **formale Semantik**, die **kombinierte textuelle und grafische Notation** sowie die **standardisierte REST-API**.

[wangxingjun778]

To further clarify: do the document paths in your retrieval corpus include any Spanish or German-language documents?

Additionally, some models—particularly smaller-sized ones—have weaker instruction-following capabilities, which can indeed lead to mixed or unexpected language outputs.

Could you please try using a larger model to see if the issue persists? If the problem still occurs, we will introduce relevant constraints and safeguards to mitigate this behavior.

[fniu]

I can confirm that with a larger model gpt-5.4, it works expected, speaking the same language.
The search directory doesn't contain any Spanish/German language documents.

GPT-5.4-mini is not a bad model, so some additional guardrail is appreciated.