Multilingual RAG Pipeline
7 messagesUsing:Gemini 2.5 Flash
Platform
Welcome to the ETH Agentic AI Platform. I'm your workspace assistant — I can help you build workflows, compare models, find tutorials, or connect with the community. What would you like to explore?
I want to build a RAG pipeline that can handle multilingual documents — German, French, and English. Which model should I use?
Router → Model Advisor
Great question. For multilingual RAG with DE/FR/EN, I'd recommend comparing these options:
**Gemini 2.5 Flash** — Excellent multilingual, fast, lowest cost ($). 1M context window handles long documents well.
**Mistral Large** — European model with strong multilingual capabilities. Good for regulatory/formal text.
**Qwen 3 235B** — Open-weight, strong multilingual. Best if you need full data sovereignty (self-hosted).
Want me to set up a side-by-side comparison with a sample of your documents?
Yes, compare Gemini Flash and Mistral on a sample regulatory document.
Evaluation Agent
model_compare(models=['gemini-flash', 'mistral-large'], task='multilingual_rag', document='sample_regulatory.pdf')Running comparison on your regulatory document sample...
| Metric | Gemini Flash | Mistral Large |
|--------|-------------|---------------|
| DE accuracy | 94.2% | 92.8% |
| FR accuracy | 93.1% | 95.4% |
| EN accuracy | 96.8% | 94.1% |
| Latency (avg) | 180ms | 400ms |
| Cost per query | CHF 0.003 | CHF 0.012 |
**Recommendation:** Gemini Flash for cost-optimized pipelines (4x cheaper). Mistral Large if French regulatory text precision is critical.
Want me to scaffold a RAG workflow with your chosen model?
Go with Gemini Flash. Set up the workflow.
Workflow Builder
create_workflow(name='Multilingual RAG Pipeline', model='gemini-flash', type='rag')✅ Workflow created: **Multilingual RAG Pipeline**
I've set up the scaffold with:
- Document ingestion (PDF/DOCX, auto-language detection)
- Chunking strategy: semantic, 512 tokens
- Embedding model: multilingual-e5-large (DE/FR/EN)
- Retriever: hybrid (semantic + BM25)
- Generator: Gemini 2.5 Flash
You can open it in the Workspace to customize, or I can walk you through the tutorial "RAG Pipeline from Scratch" step by step. What do you prefer?