Use simple prompt format

Use Mistral embeddings to match original index

Balance prompt - less restrictive while still using context

Add instruction to include math formulas from docs in LaTeX

Switch to Qwen 2.5 72B - highest parameter open source model

Switch to Llama 3.2 with chat_completion API

Use BGE-large embeddings (1024 dim) to match Pinecone index

Fix numpy float32 serialization for Pinecone

Fix embeddings with custom HFEmbeddings class

Add detailed error logging

Switch to free Zephyr model, fix inference client

Simplify app - remove PDF upload, fix dependencies

Fresh start without PDFs