Built a Nepali-language RAG over my company’s docs — what worked, what didn’t

Six weeks, internal hackathon project. Stack: Llama 3.1 8B, BGE-M3 embeddings, Qdrant.

Worked: BGE-M3 handled Nepali + English mixed docs surprisingly well, no fine-tune needed.

Didn’t work: chunk size of 512 was too small for our PDF reports, 2000 worked better. Also: do not skip evaluation — half my early “good” answers were confidently wrong.

Happy to share the eval harness in replies.