Open In Colab

Building an Advanced Fusion Retriever from Scratch#

In this tutorial, we show you how to build an advanced retriever from scratch.

Specifically, we show you how to build our QueryFusionRetriever from scratch.

This is heavily inspired from the RAG-fusion repo here: https://github.com/Raudaschl/rag-fusion.

Setup#

We load documents and build a simple vector index.

%pip install llama-index-readers-file pymupdf
%pip install llama-index-llms-openai

!pip install rank-bm25 pymupdf

import nest_asyncio

nest_asyncio.apply()

Load Documents#

!mkdir data
!wget --user-agent "Mozilla" "https://arxiv.org/pdf/2307.09288.pdf" -O "data/llama2.pdf"

If you’re opening this Notebook on colab, you will probably need to install LlamaIndex 🦙.

!pip install llama-index

from pathlib import Path
from llama_index.readers.file import PyMuPDFReader

loader = PyMuPDFReader()
documents = loader.load(file_path="./data/llama2.pdf")

Load into Vector Store#

from llama_index.core import VectorStoreIndex
from llama_index.core.node_parser import SentenceSplitter

splitter = SentenceSplitter(chunk_size=1024)
index = VectorStoreIndex.from_documents(documents, transformations=[splitter])

Define LLMs#

from llama_index.llms.openai import OpenAI

llm = OpenAI(model="gpt-3.5-turbo")

Define Advanced Retriever#

We define an advanced retriever that performs the following steps:

  1. Query generation/rewriting: generate multiple queries given the original user query

  2. Perform retrieval for each query over an ensemble of retrievers.

  3. Reranking/fusion: fuse results from all queries, and apply a reranking step to “fuse” the top relevant results!

Then in the next section we’ll plug this into our response synthesis module.

Step 1: Query Generation/Rewriting#

The first step is to generate queries from the original query to better match the query intent, and increase precision/recall of the retrieved results. For instance, we might be able to rewrite the query into smaller queries.

We can do this by prompting ChatGPT.

from llama_index.core import PromptTemplate

query_str = "How do the models developed in this work compare to open-source chat models based on the benchmarks tested?"

query_gen_prompt_str = (
    "You are a helpful assistant that generates multiple search queries based on a "
    "single input query. Generate {num_queries} search queries, one on each line, "
    "related to the following input query:\n"
    "Query: {query}\n"
    "Queries:\n"
)
query_gen_prompt = PromptTemplate(query_gen_prompt_str)

def generate_queries(llm, query_str: str, num_queries: int = 4):
    fmt_prompt = query_gen_prompt.format(
        num_queries=num_queries - 1, query=query_str
    )
    response = llm.complete(fmt_prompt)
    queries = response.text.split("\n")
    return queries

queries = generate_queries(llm, query_str, num_queries=4)

print(queries)

['1. What are the benchmarks used to evaluate open-source chat models?', '2. Can you provide a comparison between the models developed in this work and existing open-source chat models?', '3. Are there any notable differences in performance between the models developed in this work and open-source chat models based on the benchmarks tested?']

Step 2: Perform Vector Search for Each Query#

Now we run retrieval for each query. This means that we fetch the top-k most relevant results from each vector store.

NOTE: We can also have multiple retrievers. Then the total number of queries we run is NM, where N is number of retrievers and M is number of generated queries. Hence there will also be NM retrieved lists.

Here we’ll use the retriever provided from our vector store. If you want to see how to build this from scratch please see our tutorial on this.

from tqdm.asyncio import tqdm


async def run_queries(queries, retrievers):
    """Run queries against retrievers."""
    tasks = []
    for query in queries:
        for i, retriever in enumerate(retrievers):
            tasks.append(retriever.aretrieve(query))

    task_results = await tqdm.gather(*tasks)

    results_dict = {}
    for i, (query, query_result) in enumerate(zip(queries, task_results)):
        results_dict[(query, i)] = query_result

    return results_dict

# get retrievers
from llama_index.core.retrievers import BM25Retriever


## vector retriever
vector_retriever = index.as_retriever(similarity_top_k=2)

## bm25 retriever
bm25_retriever = BM25Retriever.from_defaults(
    docstore=index.docstore, similarity_top_k=2
)

results_dict = await run_queries(queries, [vector_retriever, bm25_retriever])

100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:00<00:00, 22.59it/s]

Step 3: Perform Fusion#

The next step here is to perform fusion: combining the results from several retrievers into one and re-ranking.

Note that a given node might be retrieved multiple times from different retrievers, so there needs to be a way to de-dup and rerank the node given the multiple retrievals.

We’ll show you how to perform “reciprocal rank fusion”: for each node, add up its reciprocal rank in every list where it’s retrieved.

Then reorder nodes by highest score to least.

Full paper here: https://plg.uwaterloo.ca/~gvcormac/cormacksigir09-rrf.pdf

def fuse_results(results_dict, similarity_top_k: int = 2):
    """Fuse results."""
    k = 60.0  # `k` is a parameter used to control the impact of outlier rankings.
    fused_scores = {}
    text_to_node = {}

    # compute reciprocal rank scores
    for nodes_with_scores in results_dict.values():
        for rank, node_with_score in enumerate(
            sorted(
                nodes_with_scores, key=lambda x: x.score or 0.0, reverse=True
            )
        ):
            text = node_with_score.node.get_content()
            text_to_node[text] = node_with_score
            if text not in fused_scores:
                fused_scores[text] = 0.0
            fused_scores[text] += 1.0 / (rank + k)

    # sort results
    reranked_results = dict(
        sorted(fused_scores.items(), key=lambda x: x[1], reverse=True)
    )

    # adjust node scores
    reranked_nodes: List[NodeWithScore] = []
    for text, score in reranked_results.items():
        reranked_nodes.append(text_to_node[text])
        reranked_nodes[-1].score = score

    return reranked_nodes[:similarity_top_k]

final_results = fuse_results(results_dict)

from llama_index.core.response.notebook_utils import display_source_node

for n in final_results:
    display_source_node(n, source_length=500)

Node ID: d92e53b7-1f27-4129-8d5d-dd06638b1f2d
Similarity: 0.04972677595628415
Text: Figure 12: Human evaluation results for Llama 2-Chat models compared to open- and closed-source models across ~4,000 helpfulness prompts with three raters per prompt. The largest Llama 2-Chat model is competitive with ChatGPT. Llama 2-Chat 70B model has a win rate of 36% and a tie rate of 31.5% relative to ChatGPT. Llama 2-Chat 70B model outperforms PaLM-bison chat model by a large percentage on our prompt set. More results and analysis is available in Section A.3.7. Inter-Rater Reliability (…

Node ID: 20d32df8-e16e-45fb-957a-e08175e188e8
Similarity: 0.016666666666666666
Text: Figure 1: Helpfulness human evaluation results for Llama 2-Chat compared to other open-source and closed-source models. Human raters compared model generations on ~4k prompts consisting of both single and multi-turn prompts. The 95% confidence intervals for this evaluation are between 1% and 2%. More details in Section 3.4.2. While reviewing these results, it is important to note that human evaluations can be noisy due to limitations of the prompt set, subjectivity of the review guidelines, s…

Analysis: The above code has a few straightforward components.

  1. Go through each node in each retrieved list, and add it’s reciprocal rank to the node’s ID. The node’s ID is the hash of it’s text for dedup purposes.

  2. Sort results by highest-score to lowest.

  3. Adjust node scores.

Plug into RetrieverQueryEngine#

Now we’re ready to define this as a custom retriever, and plug it into our RetrieverQueryEngine (which does retrieval and synthesis).

from llama_index.core import QueryBundle
from llama_index.core.retrievers import BaseRetriever
from typing import Any, List
from llama_index.core.schema import NodeWithScore


class FusionRetriever(BaseRetriever):
    """Ensemble retriever with fusion."""

    def __init__(
        self,
        llm,
        retrievers: List[BaseRetriever],
        similarity_top_k: int = 2,
    ) -> None:
        """Init params."""
        self._retrievers = retrievers
        self._similarity_top_k = similarity_top_k
        super().__init__()

    def _retrieve(self, query_bundle: QueryBundle) -> List[NodeWithScore]:
        """Retrieve."""
        queries = generate_queries(llm, query_str, num_queries=4)
        results = run_queries(queries, [vector_retriever, bm25_retriever])
        final_results = fuse_results(
            results_dict, similarity_top_k=self._similarity_top_k
        )

        return final_results

from llama_index.core.query_engine import RetrieverQueryEngine

fusion_retriever = FusionRetriever(
    llm, [vector_retriever, bm25_retriever], similarity_top_k=2
)

query_engine = RetrieverQueryEngine(fusion_retriever)

response = query_engine.query(query_str)

/Users/jerryliu/Programming/gpt_index/llama_index/indices/base_retriever.py:22: RuntimeWarning: coroutine 'run_queries' was never awaited
  return self._retrieve(str_or_query_bundle)
RuntimeWarning: Enable tracemalloc to get the object allocation traceback

print(str(response))

The models developed in this work, specifically the Llama 2-Chat models, are competitive with open-source chat models based on the benchmarks tested. The largest Llama 2-Chat model has a win rate of 36% and a tie rate of 31.5% relative to ChatGPT, which indicates that it performs well in comparison. Additionally, the Llama 2-Chat 70B model outperforms the PaLM-bison chat model by a large percentage on the prompt set used for evaluation. While it is important to note the limitations of the benchmarks and the subjective nature of human evaluations, the results suggest that the Llama 2-Chat models are on par with or even outperform open-source chat models in certain aspects.