Retrieval evaluation
See how the retrieval component responds to changes in the pipeline.
The retrieval component of our RAG pipeline is responsible for finding relevant documents or document chunks to feed into the generation component. In this section we'll explore how to evaluate the performance of the retrieval component of your RAG pipeline. We're checking how accurate the semantic search is, or in other words how relevant the retrieved documents are to the query.
Our retrieval component takes the incoming query and converts it into a vector or embedded representation that can be used to search for relevant documents. We then use this representation to search through a corpus of documents and retrieve the most relevant ones.
Manual evaluation using handcrafted queries
The most naive and simple way to check this would be to handcraft some queries where we know the specific documents needed to answer it. We can then check if the retrieval component is able to retrieve these documents. This is a manual evaluation process and can be time-consuming, but it's a good way to get a sense of how well the retrieval component is working. It can also be useful to target known edge cases or difficult queries to see how the retrieval component handles those known scenarios.
![](https://docs.zenml.io/~gitbook/image?url=https%3A%2F%2F3985712949-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252F-MCl0MaN3pJfPa2vq2Rb%252Fuploads%252Fgit-blob-ee4ca4ed1380b96067e58e2b285dbacb3a7e4808%252Fretrieval-eval-manual.png%3Falt%3Dmedia&width=768&dpr=4&quality=100&sign=d3a5e6a6&sv=1)
Implementing this is pretty simple - you just need to create some queries and check the retrieved documents. Having tested the basic inference of our RAG setup quite a bit, there were some clear areas where the retrieval component could be improved. I looked in our documentation to find some examples where the information could only be found in a single page and then wrote some queries that would require the retrieval component to find that page. For example, the query "How do I get going with the Label Studio integration? What are the first steps?" would require the retrieval component to find the Label Studio integration page. Some of the other examples used are:
Question | URL Ending |
---|---|
How do I get going with the Label Studio integration? What are the first steps? | stacks-and-components/component-guide/annotators/label-studio |
How can I write my own custom materializer? | user-guide/advanced-guide/data-management/handle-custom-data-types |
How do I generate embeddings as part of a RAG pipeline when using ZenML? | user-guide/llmops-guide/rag-with-zenml/embeddings-generation |
How do I use failure hooks in my ZenML pipeline? | user-guide/advanced-guide/pipelining-features/use-failure-success-hooks |
Can I deploy ZenML self-hosted with Helm? How do I do it? | deploying-zenml/zenml-self-hosted/deploy-with-helm |
For the retrieval pipeline, all we have to do is encode the query as a vector and then query the PostgreSQL database for the most similar vectors. We then check whether the URL for the document we thought must show up is actually present in the top n
results.
We include some logging so that when running the pipeline locally we can get some immediate feedback logged to the console.
This functionality can then be packaged up into a ZenML step once we're happy it does what we need:
We got a 20% failure rate on the first run of this test, which was a good sign that the retrieval component could be improved. We only had 5 test cases, so this was just a starting point. In reality, you'd want to keep adding more test cases to cover a wider range of scenarios. You'll discover these failure cases as you use the system more and more, so it's a good idea to keep a record of them and add them to your test suite.
You'd also want to examine the logs to see exactly which query failed. In our case, checking the logs in the ZenML dashboard, we find the following:
We can maybe take a look at those documents to see why they were retrieved and not the one we expected. This is a good way to iteratively improve the retrieval component.
Automated evaluation using synthetic generated queries
For a broader evaluation we can examine a larger number of queries to check the retrieval component's performance. We do this by using an LLM to generate synthetic data. In our case we take the text of each document chunk and pass it to an LLM, telling it to generate a question.
![](https://docs.zenml.io/~gitbook/image?url=https%3A%2F%2F3985712949-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252F-MCl0MaN3pJfPa2vq2Rb%252Fuploads%252Fgit-blob-b5da01f7f81d49a048cb2b3c10d64555d5aee9bd%252Fretrieval-eval-automated.png%3Falt%3Dmedia&width=768&dpr=4&quality=100&sign=e5c1337c&sv=1)
For example, given the text:
we might get the question:
If we generate questions for all of our chunks, we can then use these question-chunk pairs to evaluate the retrieval component. We pass the generated query to the retrieval component and then we check if the URL for the original document is in the top n
results.
To generate the synthetic queries we can use the following code:
As you can see, we're using litellm
again as the wrapper for the API calls. This allows us to switch between using a cloud LLM API (like OpenAI's GPT3.5 or 4) and a local LLM (like a quantized version of Mistral AI's Mixtral made available with Ollama. This has a number of advantages:
you keep your costs down by using a local model
you can iterate faster by not having to wait for API calls
you can use the same code for both local and cloud models
For some tasks you'll want to use the best model your budget can afford, but for this task of question generation we're fine using a local and slightly less capable model. Even better is that it'll be much faster to generate the questions, especially using the basic setup we have here.
To give you an indication of how long this process takes, generating 1800+ questions from an equivalent number of documentation chunks took a little over 45 minutes using the local model on a GPU-enabled machine with Ollama.
![](https://docs.zenml.io/~gitbook/image?url=https%3A%2F%2F3985712949-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252F-MCl0MaN3pJfPa2vq2Rb%252Fuploads%252Fgit-blob-67e6195c098c845c885efd7dce4bf9af6508540f%252Fhf-qa-embedding-questions.png%3Falt%3Dmedia&width=768&dpr=4&quality=100&sign=a50d8f59&sv=1)
You can view the generated dataset on the Hugging Face Hub here. This dataset contains the original document chunks, the generated questions, and the URL reference for the original document.
Once we have the generated questions, we can then pass them to the retrieval component and check the results. For convenience we load the data from the Hugging Face Hub and then pass it to the retrieval component for evaluation. We shuffle the data and select a subset of it to speed up the evaluation process, but for a more thorough evaluation you could use the entire dataset. (The best practice of keeping a separate set of data for evaluation purposes is also recommended here, though we're not doing that in this example.)
When we run this as part of the evaluation pipeline, we get a 16% failure rate which again tells us that we're doing pretty well but that there is room for improvement. As a baseline, this is a good starting point. We can then iterate on the retrieval component to improve its performance.
To take this further, there are a number of ways it might be improved:
More diverse question generation: The current question generation approach uses a single prompt to generate questions based on the document chunks. You could experiment with different prompts or techniques to generate a wider variety of questions that test the retrieval component more thoroughly. For example, you could prompt the LLM to generate questions of different types (factual, inferential, hypothetical, etc.) or difficulty levels.
Semantic similarity metrics: In addition to checking if the expected URL is retrieved, you could calculate semantic similarity scores between the query and the retrieved documents using metrics like cosine similarity. This would give you a more nuanced view of retrieval performance beyond just binary success/failure. You could track average similarity scores and use them as a target metric to improve.
Comparative evaluation: Test out different retrieval approaches (e.g. different embedding models, similarity search algorithms, etc.) and compare their performance on the same set of queries. This would help identify the strengths and weaknesses of each approach.
Error analysis: Do a deeper dive into the failure cases to understand patterns and potential areas for improvement. Are certain types of questions consistently failing? Are there common characteristics among the documents that aren't being retrieved properly? Insights from error analysis can guide targeted improvements to the retrieval component.
To wrap up, the retrieval evaluation process we've walked through - from manual spot-checking with carefully crafted queries to automated testing with synthetic question-document pairs - has provided a solid baseline understanding of our retrieval component's performance. The failure rates of 20% on our handpicked test cases and 16% on a larger sample of generated queries highlight clear room for improvement, but also validate that our semantic search is generally pointing in the right direction.
Going forward, we have a rich set of options to refine and upgrade our evaluation approach. Generating a more diverse array of test questions, leveraging semantic similarity metrics for a nuanced view beyond binary success/failure, performing comparative evaluations of different retrieval techniques, and conducting deep error analysis on failure cases - all of these avenues promise to yield valuable insights. As our RAG pipeline grows to handle more complex and wide-ranging queries, continued investment in comprehensive retrieval evaluation will be essential to ensure we're always surfacing the most relevant information.
Before we start working to improve or tweak our retrieval based on these evaluation results, let's shift gears and look at how we can evaluate the generation component of our RAG pipeline. Assessing the quality of the final answers produced by the system is equally crucial to gauging the effectiveness of our retrieval.
Retrieval is only half the story. The true test of our system is the quality of the final answers it generates by combining retrieved content with LLM intelligence. In the next section, we'll dive into a parallel evaluation process for the generation component, exploring both automated metrics and human assessment to get a well-rounded picture of our RAG pipeline's end-to-end performance. By shining a light on both halves of the RAG architecture, we'll be well-equipped to iterate and optimize our way to an ever more capable and reliable question-answering system.
Code Example
To explore the full code, visit the Complete Guide repository and for this section, particularly the eval_retrieval.py
file.
![ZenML Scarf](https://docs.zenml.io/~gitbook/image?url=https%3A%2F%2Fstatic.scarf.sh%2Fa.png%3Fx-pxid%3Df0b4f458-0a54-4fcd-aa95-d5ee424815bc&width=768&dpr=4&quality=100&sign=55709a31&sv=1)
Last updated