- Author
- Not stated
- Published
- 20 October 2025
- Kind
- Short personal blog post of practitioner lessons
- Original
- blog.abdellatif.io/production-rag-processing-5m-documents
- Subject
- Language models and training
- Themes
- Getting good
What it is
The author describes eight months of work on retrieval-augmented generation. The two projects held 9M and 4M pages. The post lists what helped and what the stack ended up being.
What it says
- The author's team built RAG for two projects, with 9M and 4M pages.
- A prototype that worked on 100 documents gave subpar results at production scale.
- The author ranks five fixes by return, starting with query generation and reranking.
- The best reranker setup was 50 chunks in and 15 out.
- The vector database moved from Azure to Pinecone to Turbopuffer.
What evidence it gives
The post is the author's account of two projects. It gives no benchmarks, figures or code beyond a stack list.
Interests and limits
The post is short and gives no named author on the page. It links the author's startup and an open-source project built from these lessons. Results are anecdotes from two deployments, with no metrics. The post does not say how rankings were measured.