Contextual Retrieval

In this notebook we will demonstrate how you can implement Anthropic’s Contextual Retrieval using LlamaIndex abstractions.

We will use:

1. Paul Graham Essay dataset.
2. Anthropic LLM for context creation for each chunk.
3. OpenAI LLM for Synthetic query generation and embedding model.
4. CohereAI Reranker.

Installation

!pip install -U llama-index llama-index-llms-anthropic llama-index-postprocessor-cohere-rerank llama-index-retrievers-bm25 stemmer

import nest_asyncio

nest_asyncio.apply()

Setup API Keys

import os

# For creating context for each chunk
os.environ["ANTHROPIC_API_KEY"] = "<YOUR ANTHROPIC API KEY>"

# For creating synthetic dataset and embedding model
os.environ["OPENAI_API_KEY"] = "<YOUR OPENAI API KEY>"

# For reranker
os.environ["COHERE_API_KEY"] = "<YOUR COHEREAI API KEY>"

Setup LLM and Embedding model

from llama_index.llms.anthropic import Anthropic

llm_anthropic = Anthropic(model="claude-3-5-sonnet-20240620")

from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.core import Settings

Settings.embed_model = OpenAIEmbedding()

Download Data

!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/paul_graham/paul_graham_essay.txt' -O './paul_graham_essay.txt'

--2024-10-01 13:00:06--  https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/paul_graham/paul_graham_essay.txt
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.108.133, 185.199.109.133, 185.199.110.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.108.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 75042 (73K) [text/plain]
Saving to: ‘./paul_graham_essay.txt’

./paul_graham_essay 100%[===================>]  73.28K  --.-KB/s    in 0.08s

2024-10-01 13:00:06 (921 KB/s) - ‘./paul_graham_essay.txt’ saved [75042/75042]

Load Data

from llama_index.core import SimpleDirectoryReader

documents = SimpleDirectoryReader(
    input_files=["./paul_graham_essay.txt"],
).load_data()

WHOLE_DOCUMENT = documents[0].text

Prompts for creating context for each chunk

We will utilize anthropic prompt caching for creating context for each chunk. If you haven’t explored our integration yet, please take a moment to review it here.

prompt_document = """<document>
{WHOLE_DOCUMENT}
</document>"""

prompt_chunk = """Here is the chunk we want to situate within the whole document
<chunk>
{CHUNK_CONTENT}
</chunk>
Please give a short succinct context to situate this chunk within the overall document for the purposes of improving search retrieval of the chunk. Answer only with the succinct context and nothing else."""

Utils

1. create_contextual_nodes - Function to create contextual nodes for a list of nodes.

2. create_embedding_retriever - Function to create an embedding retriever for a list of nodes.

3. create_bm25_retriever - Function to create a bm25 retriever for a list of nodes.

4. EmbeddingBM25RerankerRetriever - Custom retriever that uses both embedding and bm25 retrievers and reranker.

5. create_eval_dataset - Function to create a evaluation dataset from a list of nodes.

6. set_node_ids - Function to set node ids for a list of nodes.

7. retrieval_results - Function to get retrieval results for a retriever and evaluation dataset.

8. display_results - Function to display results from retrieval_results

from llama_index.retrievers.bm25 import BM25Retriever
from llama_index.core.evaluation import (
    generate_question_context_pairs,
    RetrieverEvaluator,
)
from llama_index.core.retrievers import BaseRetriever, VectorIndexRetriever
from llama_index.core.schema import NodeWithScore
from llama_index.core import VectorStoreIndex, QueryBundle
from llama_index.core.llms import ChatMessage, TextBlock

import pandas as pd
import copy
import Stemmer

from typing import List


def create_contextual_nodes(nodes_):
    """Function to create contextual nodes for a list of nodes"""
    nodes_modified = []
    for node in nodes_:
        new_node = copy.deepcopy(node)
        messages = [
            ChatMessage(role="system", content="You are helpful AI Assitant."),
            ChatMessage(
                role="user",
                content=[
                    TextBlock(
                        text=prompt_document.format(
                            WHOLE_DOCUMENT=WHOLE_DOCUMENT
                        )
                    ),
                    TextBlock(
                        text=prompt_chunk.format(CHUNK_CONTENT=node.text)
                    ),
                ],
                additional_kwargs={"cache_control": {"type": "ephemeral"}},
            ),
        ]
        new_node.metadata["context"] = str(
            llm_anthropic.chat(
                messages,
                extra_headers={"anthropic-beta": "prompt-caching-2024-07-31"},
            )
        )
        nodes_modified.append(new_node)

    return nodes_modified


def create_embedding_retriever(nodes_, similarity_top_k=2):
    """Function to create an embedding retriever for a list of nodes"""
    vector_index = VectorStoreIndex(nodes_)
    retriever = vector_index.as_retriever(similarity_top_k=similarity_top_k)
    return retriever


def create_bm25_retriever(nodes_, similarity_top_k=2):
    """Function to create a bm25 retriever for a list of nodes"""
    bm25_retriever = BM25Retriever.from_defaults(
        nodes=nodes_,
        similarity_top_k=similarity_top_k,
        stemmer=Stemmer.Stemmer("english"),
        language="english",
    )
    return bm25_retriever


def create_eval_dataset(nodes_, llm, num_questions_per_chunk=2):
    """Function to create a evaluation dataset for a list of nodes"""
    qa_dataset = generate_question_context_pairs(
        nodes_, llm=llm, num_questions_per_chunk=num_questions_per_chunk
    )
    return qa_dataset


def set_node_ids(nodes_):
    """Function to set node ids for a list of nodes"""

    # by default, the node ids are set to random uuids. To ensure same id's per run, we manually set them.
    for index, node in enumerate(nodes_):
        node.id_ = f"node_{index}"

    return nodes_


async def retrieval_results(retriever, eval_dataset):
    """Function to get retrieval results for a retriever and evaluation dataset"""

    metrics = ["hit_rate", "mrr", "precision", "recall", "ap", "ndcg"]

    retriever_evaluator = RetrieverEvaluator.from_metric_names(
        metrics, retriever=retriever
    )

    eval_results = await retriever_evaluator.aevaluate_dataset(qa_dataset)

    return eval_results


def display_results(name, eval_results):
    """Display results from evaluate."""

    metrics = ["hit_rate", "mrr", "precision", "recall", "ap", "ndcg"]

    metric_dicts = []
    for eval_result in eval_results:
        metric_dict = eval_result.metric_vals_dict
        metric_dicts.append(metric_dict)

    full_df = pd.DataFrame(metric_dicts)

    columns = {
        "retrievers": [name],
        **{k: [full_df[k].mean()] for k in metrics},
    }

    metric_df = pd.DataFrame(columns)

    return metric_df


class EmbeddingBM25RerankerRetriever(BaseRetriever):
    """Custom retriever that uses both embedding and bm25 retrievers and reranker"""

    def __init__(
        self,
        vector_retriever: VectorIndexRetriever,
        bm25_retriever: BM25Retriever,
        reranker: CohereRerank,
    ) -> None:
        """Init params."""

        self._vector_retriever = vector_retriever
        self.bm25_retriever = bm25_retriever
        self.reranker = reranker

        super().__init__()

    def _retrieve(self, query_bundle: QueryBundle) -> List[NodeWithScore]:
        """Retrieve nodes given query."""

        vector_nodes = self._vector_retriever.retrieve(query_bundle)
        bm25_nodes = self.bm25_retriever.retrieve(query_bundle)

        vector_nodes.extend(bm25_nodes)

        retrieved_nodes = self.reranker.postprocess_nodes(
            vector_nodes, query_bundle
        )

        return retrieved_nodes

Create Nodes

from llama_index.core.node_parser import SentenceSplitter

node_parser = SentenceSplitter(chunk_size=1024, chunk_overlap=200)

nodes = node_parser.get_nodes_from_documents(documents, show_progress=False)

Set node ids

Useful to have consistent result comparison for nodes with and without contextual text.

# set node ids

nodes = set_node_ids(nodes)

nodes[0].metadata

{'file_path': 'paul_graham_essay.txt',
 'file_name': 'paul_graham_essay.txt',
 'file_type': 'text/plain',
 'file_size': 75042,
 'creation_date': '2024-10-01',
 'last_modified_date': '2024-10-01'}

Create contextual nodes

nodes_contextual = create_contextual_nodes(nodes)

nodes[0].metadata, nodes_contextual[0].metadata

({'file_path': 'paul_graham_essay.txt',
  'file_name': 'paul_graham_essay.txt',
  'file_type': 'text/plain',
  'file_size': 75042,
  'creation_date': '2024-10-01',
  'last_modified_date': '2024-10-01'},
 {'file_path': 'paul_graham_essay.txt',
  'file_name': 'paul_graham_essay.txt',
  'file_type': 'text/plain',
  'file_size': 75042,
  'creation_date': '2024-10-01',
  'last_modified_date': '2024-10-01',
  'context': 'assistant: This chunk is the opening section of Paul Graham\'s essay "What I Worked On," describing his early experiences with programming and writing as a teenager, his initial interest in philosophy in college, and his subsequent shift to studying artificial intelligence in the mid-1980s.'})

Set similarity_top_k

similarity_top_k = 3

Set CohereReranker

from llama_index.postprocessor.cohere_rerank import CohereRerank

cohere_rerank = CohereRerank(
    api_key=os.environ["COHERE_API_KEY"], top_n=similarity_top_k
)

Create retrievers.

1. Embedding based retriever.
2. BM25 based retriever.
3. Embedding + BM25 + Cohere reranker retriever.

embedding_retriever = create_embedding_retriever(
    nodes, similarity_top_k=similarity_top_k
)
bm25_retriever = create_bm25_retriever(
    nodes, similarity_top_k=similarity_top_k
)
embedding_bm25_retriever_rerank = EmbeddingBM25RerankerRetriever(
    embedding_retriever, bm25_retriever, reranker=cohere_rerank
)

DEBUG:bm25s:Building index from IDs objects

Create retrievers using contextual nodes.

contextual_embedding_retriever = create_embedding_retriever(
    nodes_contextual, similarity_top_k=similarity_top_k
)
contextual_bm25_retriever = create_bm25_retriever(
    nodes_contextual, similarity_top_k=similarity_top_k
)
contextual_embedding_bm25_retriever_rerank = EmbeddingBM25RerankerRetriever(
    contextual_embedding_retriever,
    contextual_bm25_retriever,
    reranker=cohere_rerank,
)

DEBUG:bm25s:Building index from IDs objects

Create Synthetic query dataset

from llama_index.llms.openai import OpenAI

llm = OpenAI(model="gpt-4")

qa_dataset = create_eval_dataset(nodes, llm=llm, num_questions_per_chunk=2)

100%|██████████| 21/21 [02:59<00:00,  8.53s/it]

list(qa_dataset.queries.values())[1]

"The author initially intended to study philosophy in college but later switched to AI. Discuss the reasons behind this shift in interest and how specific influences like Heinlein's novel and Winograd's SHRDLU contributed to his decision."

Evaluate retrievers with and without contextual nodes

embedding_retriever_results = await retrieval_results(
    embedding_retriever, qa_dataset
)
bm25_retriever_results = await retrieval_results(bm25_retriever, qa_dataset)
embedding_bm25_retriever_rerank_results = await retrieval_results(
    embedding_bm25_retriever_rerank, qa_dataset
)

contextual_embedding_retriever_results = await retrieval_results(
    contextual_embedding_retriever, qa_dataset
)
contextual_bm25_retriever_results = await retrieval_results(
    contextual_bm25_retriever, qa_dataset
)
contextual_embedding_bm25_retriever_rerank_results = await retrieval_results(
    contextual_embedding_bm25_retriever_rerank, qa_dataset
)

Display results

Without Context

pd.concat(
    [
        display_results("Embedding Retriever", embedding_retriever_results),
        display_results("BM25 Retriever", bm25_retriever_results),
        display_results(
            "Embedding + BM25 Retriever + Reranker",
            embedding_bm25_retriever_rerank_results,
        ),
    ],
    ignore_index=True,
    axis=0,
)

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

	retrievers	hit_rate	mrr	precision	recall	ap	ndcg
0	Embedding Retriever	0.857143	0.726190	0.285714	0.857143	0.726190	0.356613
1	BM25 Retriever	0.904762	0.777778	0.301587	0.904762	0.777778	0.380157
2	Embedding + BM25 Retriever + Reranker	0.952381	0.865079	0.456349	0.952381	0.865079	0.530172

<script>
  const buttonEl =
    document.querySelector('#df-d3120688-09ee-4895-b98e-b26fe4dab8b3 button.colab-df-convert');
  buttonEl.style.display =
    google.colab.kernel.accessAllowed ? 'block' : 'none';

  async function convertToInteractive(key) {
    const element = document.querySelector('#df-d3120688-09ee-4895-b98e-b26fe4dab8b3');
    const dataTable =
      await google.colab.kernel.invokeFunction('convertToInteractive',
                                                [key], {});
    if (!dataTable) return;

    const docLinkHtml = 'Like what you see? Visit the ' +
      '<a target="_blank" href=https://colab.research.google.com/notebooks/data_table.ipynb>data table notebook</a>'
      + ' to learn more about interactive tables.';
    element.innerHTML = '';
    dataTable['output_type'] = 'display_data';
    await google.colab.output.renderOutput(dataTable, element);
    const docLink = document.createElement('div');
    docLink.innerHTML = docLinkHtml;
    element.appendChild(docLink);
  }
</script>

<svg xmlns=“http://www.w3.org/2000/svg” height=“24px”viewBox=“0 0 24 24” width=“24px”>

</div>

With Context

pd.concat(
    [
        display_results(
            "Contextual Embedding Retriever",
            contextual_embedding_retriever_results,
        ),
        display_results(
            "Contextual BM25 Retriever", contextual_bm25_retriever_results
        ),
        display_results(
            "Contextual Embedding + Contextual BM25 Retriever + Reranker",
            contextual_embedding_bm25_retriever_rerank_results,
        ),
    ],
    ignore_index=True,
    axis=0,
)

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

	retrievers	hit_rate	mrr	precision	recall	ap	ndcg
0	Contextual Embedding Retriever	0.928571	0.746032	0.309524	0.928571	0.746032	0.372175
1	Contextual BM25 Retriever	0.952381	0.829365	0.317460	0.952381	0.829365	0.403967
2	Contextual Embedding + Contextual BM25 Retriev...	0.976190	0.900794	0.476190	0.976190	0.900794	0.555746

<script>
  const buttonEl =
    document.querySelector('#df-9344898e-8549-43d4-9f09-e83f6834b765 button.colab-df-convert');
  buttonEl.style.display =
    google.colab.kernel.accessAllowed ? 'block' : 'none';

  async function convertToInteractive(key) {
    const element = document.querySelector('#df-9344898e-8549-43d4-9f09-e83f6834b765');
    const dataTable =
      await google.colab.kernel.invokeFunction('convertToInteractive',
                                                [key], {});
    if (!dataTable) return;

    const docLinkHtml = 'Like what you see? Visit the ' +
      '<a target="_blank" href=https://colab.research.google.com/notebooks/data_table.ipynb>data table notebook</a>'
      + ' to learn more about interactive tables.';
    element.innerHTML = '';
    dataTable['output_type'] = 'display_data';
    await google.colab.output.renderOutput(dataTable, element);
    const docLink = document.createElement('div');
    docLink.innerHTML = docLinkHtml;
    element.appendChild(docLink);
  }
</script>

<svg xmlns=“http://www.w3.org/2000/svg” height=“24px”viewBox=“0 0 24 24” width=“24px”>

</div>

Observation:

We observed improved metrics with contextual retrieval; however, our experiments showed that much depends on the queries, chunk size, chunk overlap, and other variables. Therefore, it’s essential to experiment to optimize the benefits of this technique.