double-layered-rag-benchmarks

Small benchmark test results

This small benchmark was conceived to test the solidity of the double-layered RAG approach implemented inside qdurllm as a retrieval technique.

RAG workflow

The RAG workflow goes like this:

• All text batches, obtained by chunking URL contents, are first encoded by All-MiniLM-L6-v2 and uploaded to a permanent Qdrant collection
• The first retrieval call extract the 10 closest matches to the query in terms of cosine distance, based on All-MiniLM-L6-v2 search
• The 10 best hits get re-encoded by sentence-t5-base and uploaded to a non-permanent Qdrant collection (gets deleted at every round after use)
• sentence-t5-base performs a second retrieval call extracting the best match to the original query, and this best match gets returned

First test

Benchmark is based on the content of 4 web pages:

• https://www.technologyreview.com/2023/12/05/1084417/ais-carbon-footprint-is-bigger-than-you-think/
• https://semiengineering.com/ai-power-consumption-exploding/
• https://www.piie.com/blogs/realtime-economics/2024/ais-carbon-footprint-appears-likely-be-alarming
• https://news.climate.columbia.edu/2023/06/09/ais-growing-carbon-footprint/

The content of these URLs was chunked up and uploaded to Qdrant collections, and at the same time smaller portions of each chunk (encompassing 10-25% of the text) were used for querying, and the retrieved results compared with the original full text.

First results

The correct/total retrievals ratio for the only All-MiniLM-L6-v2 is 81.54%, whereas the correct/total retrievals ratio for the previously described double-layered All-MiniLM-L6-v2 + sentence-t5-base goes up 93.85%, equalling the one of sentence-t5-base alone. Following a double-layered approach with switched roles for the two encoders yields a correct/total retrievals ratio of 84.62%.

The advantage of this technique is that it does not require that all the chunks of text are encoded in 768-dimensional vectors (as would happen if we adopted sentence-t5-base alone), but this step is done dynamically at each vector call. As you can see, it also definitely improves the performance of the sole All-MiniLM-L6-v2 by little more than 12%.

The disadvantage is in the execution time: on a 8GB RAM-12 cores Windows 10 laptop, double-layered RAG takes an average of 8.39 s, against the 0.23 s of the sole sentence-t5-base.

Second test

The second benchmark is based on dataset, available on HuggingFace. It is a Q&A dataset based on a set of 358 answers (used as content to retrieve) and questions (used as retrieval queries).

Second results

• Avg time for All-MiniLM-L6-v2: 0.13889479704117508 +/- 0.018010675079187972
• Avg time for sentence-t5-base: 0.3546350625123871 +/- 0.1378480017367839
• Avg time for All-MiniLM-L6-v2 + sentence-t5-base: 10.722357098306164 +/- 1.2552639024596886
• Avg time for sentence-t5-base + All-MiniLM-L6-v2: 2.722710152020615 +/- 0.30703480688834284
• Correct/Total retrievals for All-MiniLM-L6-v2: 0.41853932584269665
• Correct/Total retrievals for sentence-t5-base: 0.5084269662921348
• Correct/Total retrievals for All-MiniLM-L6-v2 + sentence-t5-base: 0.5028089887640449
• Correct/Total retrievals for sentence-t5-base + All-MiniLM-L6-v2: 0.42134831460674155

Code availability

The benchmark test code is available here for the first test and here for the second one.

Contributions

If you happen to have time and a powerful hardware, you can carry on vaster tests using the script referenced before: it would be great!