引入上下文检索（Contextual Retrieval）

For an AI model to be useful in specific contexts, it often needs access to background knowledge.

For example, customer support chatbots need knowledge about the specific business they're being used for, and legal analyst bots need to know about a vast array of past cases.

Developers typically enhance an AI model's knowledge using Retrieval-Augmented Generation (RAG).

RAG is a method that retrieves relevant information from a knowledge base and appends it to the user's prompt, significantly enhancing the model's response.

The problem is that traditional RAG solutions remove context when encoding information, which often results in the system failing to retrieve the relevant information from the knowledge base.

In this post, we outline a method that dramatically improves the retrieval step in RAG.

The method is called “Contextual Retrieval” and uses two sub-techniques: Contextual Embeddings and Contextual BM25.

This method can reduce the number of failed retrievals by 49% and, when combined with reranking, by 67%.

These represent significant improvements in retrieval accuracy, which directly translates to better performance in downstream tasks.

You can easily deploy your own Contextual Retrieval solution with Claude with our cookbook.

Sometimes the simplest solution is the best.

If your knowledge base is smaller than 200,000 tokens (about 500 pages of material), you can just include the entire knowledge base in the prompt that you give the model, with no need for RAG or similar methods.

A few weeks ago, we released prompt caching for Claude, which makes this approach significantly faster and more cost-effective.

Developers can now cache frequently used prompts between API calls, reducing latency by > 2x and costs by up to 90% (you can see how it works by reading our prompt caching cookbook).

However, as your knowledge base grows, you'll need a more scalable solution.

That’s where Contextual Retrieval comes in.

For larger knowledge bases that don't fit within the context window, RAG is the typical solution.

RAG works by preprocessing a knowledge base using the following steps:

At runtime, when a user inputs a query to the model, the vector database is used to find the most relevant chunks based on semantic similarity to the query.

Then, the most relevant chunks are added to the prompt sent to the generative model.

While embedding models excel at capturing semantic relationships, they can miss crucial exact matches.

Fortunately, there’s an older technique that can assist in these situations.

BM25 (Best Matching 25) is a ranking function that uses lexical matching to find precise word or phrase matches.

It's particularly effective for queries that include unique identifiers or technical terms.

BM25 works by building upon the TF-IDF (Term Frequency-Inverse Document Frequency) concept.

TF-IDF measures how important a word is to a document in a collection.

BM25 refines this by considering document length and applying a saturation function to term frequency, which helps prevent common words from dominating the results.

Here’s how BM25 can succeed where semantic embeddings fail: Suppose a user queries "Error code TS-999" in a technical support database.

An embedding model might find content about error codes in general, but could miss the exact "TS-999" match.

BM25 looks for this specific text string to identify the relevant documentation.

RAG solutions can more accurately retrieve the most applicable chunks by combining the embeddings and BM25 techniques using the following steps:

By leveraging both BM25 and embedding models, traditional RAG systems can provide more comprehensive and accurate results, balancing precise term matching with broader semantic understanding.

This approach allows you to cost-effectively scale to enormous knowledge bases, far beyond what could fit in a single prompt.

But these traditional RAG systems have a significant limitation: they often destroy context.

In traditional RAG, documents are typically split into smaller chunks for efficient retrieval.

While this approach works well for many applications, it can lead to problems when individual chunks lack sufficient context.

For example, imagine you had a collection of financial information (say, U.S. SEC filings) embedded in your knowledge base, and you received the following question: "What was the revenue growth for ACME Corp in Q2 2023?"

A relevant chunk might contain the text: "The company's revenue grew by 3% over the previous quarter."

However, this chunk on its own doesn't specify which company it's referring to or the relevant time period, making it difficult to retrieve the right information or use the information effectively.

Contextual Retrieval solves this problem by prepending chunk-specific explanatory context to each chunk before embedding (“Contextual Embeddings”) and creating the BM25 index (“Contextual BM25”).

Let’s return to our SEC filings collection example.

Here's an example of how a chunk might be transformed:

It is worth noting that other approaches to using context to improve retrieval have been proposed in the past.

Other proposals include: adding generic document summaries to chunks (we experimented and saw very limited gains), hypothetical document embedding, and summary-based indexing (we evaluated and saw low performance).

These methods differ from what is proposed in this post.

Of course, it would be far too much work to manually annotate the thousands or even millions of chunks in a knowledge base.

To implement Contextual Retrieval, we turn to Claude.

We’ve written a prompt that instructs the model to provide concise, chunk-specific context that explains the chunk using the context of the overall document.

We used the following Claude 3 Haiku prompt to generate context for each chunk:

The resulting contextual text, usually 50-100 tokens, is prepended to the chunk before embedding it and before creating the BM25 index.

Here’s what the preprocessing flow looks like in practice:

If you’re interested in using Contextual Retrieval, you can get started with our cookbook.

Contextual Retrieval is uniquely possible at low cost with Claude, thanks to the special prompt caching feature we mentioned above.

With prompt caching, you don’t need to pass in the reference document for every chunk.

You simply load the document into the cache once and then reference the previously cached content.

Assuming 800 token chunks, 8k token documents, 50 token context instructions, and 100 tokens of context per chunk, the one-time cost to generate contextualized chunks is $1.02 per million document tokens.

We experimented across various knowledge domains (codebases, fiction, ArXiv papers, Science Papers), embedding models, retrieval strategies, and evaluation metrics.

We’ve included a few examples of the questions and answers we used for each domain in Appendix II.

The graphs below show the average performance across all knowledge domains with the top-performing embedding configuration (Gemini Text 004) and retrieving the top-20-chunks.

We use 1 minus recall@20 as our evaluation metric, which measures the percentage of relevant documents that fail to be retrieved within the top 20 chunks.

You can see the full results in the appendix - contextualizing improves performance in every embedding-source combination we evaluated.

Our experiments showed that:

When implementing Contextual Retrieval, there are a few considerations to keep in mind:

Always run evals: Response generation may be improved by passing it the contextualized chunk and distinguishing between what is context and what is the chunk.

In a final step, we can combine Contextual Retrieval with another technique to give even more performance improvements.

In traditional RAG, the AI system searches its knowledge base to find the potentially relevant information chunks.

With large knowledge bases, this initial retrieval often returns a lot of chunks—sometimes hundreds—of varying relevance and importance.

Reranking is a commonly used filtering technique to ensure that only the most relevant chunks are passed to the model.

Reranking provides better responses and reduces cost and latency because the model is processing less information.

The key steps are:

There are several reranking models on the market.

We ran our tests with the Cohere reranker.

Voyage also offers a reranker, though we did not have time to test it.

Our experiments showed that, across various domains, adding a reranking step further optimizes retrieval.

Specifically, we found that Reranked Contextual Embedding and Contextual BM25 reduced the top-20-chunk retrieval failure rate by 67% (5.7% → 1.9%).

One important consideration with reranking is the impact on latency and cost, especially when reranking a large number of chunks.

Because reranking adds an extra step at runtime, it inevitably adds a small amount of latency, even though the reranker scores all the chunks in parallel.

There is an inherent trade-off between reranking more chunks for better performance vs. reranking fewer for lower latency and cost.

We recommend experimenting with different settings on your specific use case to find the right balance.

We ran a large number of tests, comparing different combinations of all the techniques described above (embedding model, use of BM25, use of contextual retrieval, use of a reranker, and total # of top-K results retrieved), all across a variety of different dataset types.

Here’s a summary of what we found:

We encourage all developers working with knowledge bases to use our cookbook to experiment with these approaches to unlock new levels of performance.

Below is a breakdown of results across datasets, embedding providers, use of BM25 in addition to embeddings, use of contextual retrieval, and use of reranking for Retrievals @ 20.

See Appendix II for the breakdowns for Retrievals @ 10 and @ 5 as well as example questions and answers for each dataset.

Research and writing by Daniel Ford.

Thanks to Orowa Sikder, Gautam Mittal, and Kenneth Lien for critical feedback, Samuel Flamini for implementing the cookbooks, Lauren Polansky for project coordination and Alex Albert, Susan Payne, Stuart Ritchie, and Brad Abrams for shaping this blog post.