LlamaIndex的5种高级RAG方法

LlamaPacks: Building RAG in Fewer Lines of Code.

Comparison conducted among the following advanced retrieval with Llamapacks.

• Baseline: Naive query engine with Llamaindex

• Pack 1: Hybrid Fusion Retriever Pack?[Notebook]?[Source]

Built on top of?QueryFusionRetriever. Generates multiple queries from the input question and then ensembles vector and bm25 retrievers using fusion(vector + keyword search + reciprocal_rerank_fusion)

• Pack 2: Query Rewriting Retriever Pack?[Notebook]?[Source]

Built on top of?QueryFusionRetriever. Rewrite(Generates) multiple queries from the input question, retrieve with vector search and rerank the results with reciprocal_rerank_fusion.

• Pack 3: AutoMergingRetrieverPack pack?[Notebook]?Source

Built on top of?AutoMergingRetriever. It loads a document, builds a hierarchical node graph (with bigger parent nodes and smaller child nodes) and then check if enough children nodes of a parent node have been retrieved and merge and replace with that parent.

• Pack 4: Small-to-big Retrieval Pack?[Notebook]?[Source]

Built on top of?RecursiveRetriever. Given input documents, and an initial set of "parent" chunks, subdivide each chunk further into "child" chunks. Link each child chunk to its parent chunk, and index the child chunks.

• Pack 5: Sentence Window Retrieval Pack?[Notebook]?[Source]

Built on top of?SentenceWindowNodeParser. It loads a document, split a document into Nodes, with each node being a sentence. Each node contains a window from the surrounding sentences in the metadata.

MODEL = "gpt-3.5-turbo"So, which one will perform best? My answer is: "It depends! Different retrieval tasks and datasets have varying requirements, and there's no one-size-fits-all solution. "