Skip to main content
Ryan Orban

Ryan Orban

Subject
7 entries

Retrieval

Bookmarks

  1. Dense-X-Retrieval: Proposition-Level RAG

    Dense-X-Retrieval is a LlamaIndex pack implementing proposition-level retrieval — splitting documents into atomic factual propositions rather than chunks, then retrieving at proposition granularity. Improves precision by matching query semantics at a finer level than paragraph chunks.

  2. RAG at Planet Scale

    Arcus describes their multi-tiered RAG approach for handling massive external data corpora at planet scale — one of the largest RAG deployments of 2023. The key innovation is tiered retrieval that narrows the candidate pool progressively rather than searching the full index directly.

  3. LlamaIndex Retrieval and Chunk Evaluation Notebook

    A LlamaIndex Google Colab notebook for evaluating retrieval quality and chunk size in RAG pipelines — demonstrating how to measure retrieval hit rate and MRR across different chunk sizes. Practical tooling for the underappreciated problem of RAG evaluation.

  4. RAG Is More Than Just Embedding Search

    Jason Liu's influential post arguing that RAG systems need more than vector similarity search — covering query understanding, routing, reranking, and structured extraction as the layers that separate working RAG from production-grade RAG. Written for the Instructor library blog.

  5. Generate Rather Than Retrieve: Large Language Models Are Strong Context Generators

    Yu et al. (2022) show that prompting an LLM to generate its own background context before answering a question (GenRead) outperforms retrieval-based approaches on several knowledge-intensive NLP benchmarks. The result challenges the assumption that retrieval is necessary for grounding LLM outputs.

  6. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Lewis et al. (Facebook AI, 2020) introduce Retrieval-Augmented Generation, a hybrid architecture that combines dense passage retrieval with seq2seq generation to ground language model outputs in a non-parametric knowledge store. RAG defined the template that most production knowledge-grounded LLM systems follow today.

  7. TART: Task-Aware Retrieval with Instructions

    TART (Task-Aware Retrieval with Instructions) introduces BERRI, a dataset of ~40 retrieval tasks annotated with human-written task instructions, and trains a multi-task retrieval system that adapts its behavior based on explicit instructions. TART outperforms much larger models on BEIR by understanding the user's intent rather than just matching query-document similarity.

All bookmarks