Building Ask Leya: Lessons Learned from Local ONNX Embeddings, Prisma Migrations, and Express Async Handlers

Description and use

Built a chat widget using local ONNX embeddings for multilingual retrieval, streamlined Prisma migrations, and robust async error handling in Express.

Problem

Need a chat widget that answers from a business’s own content, supports multiple languages locally, handles database migrations with unreleased modules, and properly forwards async errors in Express.

Outcome

The project achieved a multilingual, locally hosted chat widget with efficient embeddings, streamlined migrations, and robust async error handling, reducing external dependencies and improving response reliability.

Workflow

• 1. Load paraphrase-multilingual-MiniLM-L12-v2 ONNX model in-process for vector embeddings.
• 2. Embed user query and match against English knowledge base using pgvector.
• 3. Return answer in user language via chat model.
• 4. Use prisma migrate diff to filter additive statements for unreleased tables.
• 5. Add Express error-handling middleware to catch async rejections and send responses.

Evidence

Cross-language retrieval is local, not the LLM Retrieval uses paraphrase-multilingual-MiniLM-L12-v2 through ONNX, in-process — 384 dims, trained across 50+ languages.

Source

Express 4 doesn't forward async rejections

Source