← Back to Blogs
Skip to main content

Dense X Retrieval: What Retrieval Granularity Should We Use?

· 2 min read
Zain Hasan

A preview of the paper

❓What text chunk size should we use in our RAG workflows? How does chunk size impact retrieval recall? Are bigger chunks better? Smaller chunks but keep more top-k?

📜The new paper from Tencent and Carnegie Mellon(https://arxiv.org/abs/2312.06648) asked:

  1. What chunk size is best to segment and index a vector database like @weaviate_io ?
  2. How does chunk size impact generalization for passage retrieval and accuracy for QA RAG tasks?

⏩In Short: They found that instead of using 100-word passage or sentence-level chunking it's best create Propositions - concise, distinct and self-contained expressions of factoids.

Propositions are generated by a finetuned LLM - which takes in paragraphs as input and is instructed to generate propositions.(blue in the image)

Going to try this out with the current @weaviate_io workflows.

📑The details:

  1. QnA RAG Improvements: +5.9, +7.8,+5.8, +4.9, +5.9, and +6.9 EM@100(exact match using 100 words) for SimCSE, Contriever, DPR, ANCE, TAS-B, and GTR.

  2. Passage Retrieval Perf. : Improvement of Recall@20 is +10.1% and +2.2% for unsupervised and supervised retrievers resp.

  3. Propositions have the following properties: a. unique: a distinct piece of meaning in text b. atomic: cannot be further split into separate propositions c. self-contained: includes all the necessary context

  4. The paragraph-to-proposition generating LLM (a FlanT5-large model) is finetuned using a 42k passage dataset that has been atomized into propositions using GPT-4 - ie. the process is automatable.

  5. Supervised retrievers show less improvements with Propositions b/c these retrievers are trained on query-passage pairs.

  6. Unsupervised retrieval by proposition demonstrates a clear advantage - 17-25% Recall@5 relative improvement on EntityQuestions with DPR and ANCE.

  7. Works better for rare concepts: Retrieving by proposition much more advantageous for questions targeting less common entities.

  8. The RAG (retrieve-then-read) task uses a T5-large size UnifiedQA-v2 as the reader model.

  9. Proposition chunks outperform passage chunks for QnA most in the range of 100-200 words = ~10 propositions = ~5 sentences = ~2 passages.

🔗 arXiv Link

📜 Download paper

Ready to start building?​

Check out the Quickstart tutorial, or build amazing apps with a free trial of Weaviate Cloud (WCD).

GitHub

Forum

Slack

X (Twitter)

Don't want to miss another blog post?

Sign up for our bi-weekly newsletter to stay updated!


By submitting, I agree to the Terms of Service and Privacy Policy.