Building with agents 0/15

Building with agents

Retrieval and RAG

Retrieval-augmented generation embeds documents, finds the passages nearest a question in a vector index and gives them to the model to answer from. It is the usual way to ground a model in private or recent data without retraining it.

Reading

WhiteBox lessons3

Papers and essays3

  1. Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs(opens in a new tab)

    Malkov and Yashunin 2020 IEEE Transactions on Pattern Analysis and Machine Intelligence 60 min

    HNSW: approximate nearest-neighbor search over stacked proximity graphs, where each element appears in a random number of layers and a search starts in the sparse top layer and descends, giving logarithmic scaling.

  2. Dense Passage Retrieval for Open-Domain Question Answering(opens in a new tab)

    Karpukhin et al. 2020 EMNLP 2020 45 min

    Dense passage retrieval: two BERT encoders, one for questions and one for passages, trained with in-batch negatives so relevant pairs score high by dot product. It beat a strong BM25 system by 9 to 19 points in top-20 retrieval accuracy.

  3. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks(opens in a new tab)

    Lewis et al. 2020 NeurIPS 2020 50 min

    Retrieval-augmented generation: a BART generator conditioned on Wikipedia passages found by a DPR retriever, fine-tuned together, so knowledge can be updated by swapping the index without retraining.

Videos1

Building with LLMs Various speakers 1 video