|
Download docs/RAG_PROTOCOL.md from Angshul/SparseGeometricRAG: direct link, hf CLI and curl.
- Browser
- Download file 2.15 kB
-
https://huggingface.co/datasets/Angshul/SparseGeometricRAG/resolve/main/docs/RAG_PROTOCOL.md
- Command line
-
hf download hf://datasets/Angshul/SparseGeometricRAG/docs/RAG_PROTOCOL.md
-
curl -L -o RAG_PROTOCOL.md https://huggingface.co/datasets/Angshul/SparseGeometricRAG/resolve/main/docs/RAG_PROTOCOL.md
2.15 kB
| # Practical RAG evaluation protocol: measure the first ten chunks | |
| ## Why K=10 is the deployment target | |
| A RAG system does not consume an abstract recall curve. It consumes a small context set. Chunks ranked at 100, 500, or 1000 are normally never passed to the generator. Optimizing deep recall can therefore reward a retriever for spending CPU, memory bandwidth, and storage on results that have zero downstream utility. | |
| For this repository, the primary protocol is deliberately strict: | |
| - retrieve a ranked **top 10**; | |
| - report **nDCG@10, MRR@10, Precision@10, Recall@10, Hit@10**; | |
| - measure **CPU median latency, p95 latency, and QPS**; | |
| - report the number of routed candidates and the chunk shortlist size `P`; | |
| - use deep-recall metrics only as diagnostics, never as the main optimization objective. | |
| This is the only protocol in the repository used to decide whether a change helps practical RAG. | |
| ## Why shortlist size P is a RAG parameter | |
| The historical system used `P=2000` because it was selected for Recall@100. Once the objective became top-10 RAG quality, the relevant question changed to: | |
| > How many retrieved geometric representations should be exposed to chunk-level scoring before choosing ten chunks? | |
| The current sweep is therefore `P in {25, 50, 100, 200, 500}` plus the historical `P=2000` reference. TREC-COVID has a large route and peaks sharply at `P=100`; SciFact has a tiny route (~157 candidates/query) and improves as the artificial pruning is removed. The lesson is not that 100 is universal. The lesson is that **the shortlist must be evaluated for the top-10 deployment objective rather than inherited from a deep-recall benchmark.** | |
| ## Timing discipline | |
| Speed is a first-class result. | |
| 1. CPU is the principal deployment regime. | |
| 2. Warm the index before timing. | |
| 3. Record median and p95, not only an average. | |
| 4. Report query representation time separately when a baseline uses a neural encoder. | |
| 5. Never compare our end-to-end latency with an ANN-only latency that silently excludes dense query encoding. | |
| 6. Mark diagnostic Python implementations as diagnostic; do not present them as optimized production latency. | |