YACHAY

Problem

Answer questions over a document corpus accurately, without relying on a large budget or expensive managed infrastructure.

Constraints

• Budget: prioritize free or local tooling wherever possible.
• Accuracy: avoid model hallucinations, citing real sources from the corpus.

Architecture

RAG agent built with OCI Generative AI + LlamaIndex + ChromaDB. Embeddings via BAAI/bge-m3, with meta.llama-3.3-70b-instruct as the LLM.

Decisions

• bge-m3 for being a multilingual embeddings model, relevant for a Spanish-language corpus.
• ChromaDB for being local and free, avoiding the cost of a managed vector database.

My role

Individual project, built as part of an Alura challenge.

Result

Personal project, built for the Alura challenge. No production data yet — it’s a learning project focused on validating the RAG architecture.