Hymalaia can be configured to support multiple languages beyond English. This guide walks you through how it works and how to set it up.

​
How it works

Hymalaia relies on two core components that assume English by default:

  1. Vector Search and Reranking
    These components use embedding models to retrieve relevant documents for the LLM.

  2. LLM Prompts
    These are used to guide the LLM in how to respond.

To support multilingual scenarios:

  • Swap out English-first embedding/reranking models for multilingual models.
  • Apply query expansion to rephrase the user query into target languages.
  • Provide LLM instructions to respond in the same language as the user query.

⚠️ Note
The built-in LLM prompts are still in English. For fully non-English use, consider translating all prompts directly into your target language.

​
Configuration

Hymalaia supports multilingual setup entirely through environment variables. For Docker Compose, create a .env file inside the hymalaia/deployment/docker_compose folder.

Here’s an example configuration for English and French:

# Rephrase user query in multiple languages
MULTILINGUAL_QUERY_EXPANSION="English, French"

# Use a multilingual embedding model
DOCUMENT_ENCODER_MODEL="intfloat/multilingual-e5-small"

# Prefixes used by the embedding model
ASYM_QUERY_PREFIX="query: "
ASYM_PASSAGE_PREFIX="passage: "

# Normalize embeddings (model dependent)
NORMALIZE_EMBEDDINGS="True"

# Disable LLM chunk filtering for better multilingual support
DISABLE_LLM_CHUNK_FILTER="True"

# Disable reranking (English-first models are not optimal for multilingual use)
ENABLE_RERANKING_ASYNC_FLOW="False"
ENABLE_RERANKING_REAL_TIME_FLOW="False"

# Enable fine-grained mini-chunking for better recall
ENABLE_MINI_CHUNK="True"

# Use a stronger model for better multilingual understanding
GEN_AI_MODEL_VERSION="gpt-4"

📝 An up-to-date template for multilingual settings is also available in the codebase under hymalaia/deployment/docker_compose.

​
Recommendations

  • For full translation, localize all prompt templates in your target language.

  • Use multilingual embedding models from trusted sources (e.g. Hugging Face).

  • Prefer LLMs with multilingual support (e.g. GPT-4, Claude, etc.).