This repo contains scripts and tools for evaluating a chat app that uses the RAG architecture. There are many parameters that affect the quality and style of answers generated by the chat app, such as the system prompt, search parameters, and GPT model parameters.
Whenever you are making changes to a RAG chat with the goal of improving the answers, you should evaluate the results. This repository offers tools to make it easier to run evaluations, plus examples of evaluations that we've run on our sample chat app.
📺 Watch a video overview of this repo
Table of contents:
- Setting up this project
- Deploying a GPT-4 model
- Generating ground truth data
- Running an evaluation
- Viewing the results
- Measuring app's ability to say "I don't know"
If you open this project in a Dev Container or GitHub Codespaces, it will automatically set up the environment for you. If not, then follow these steps:
-
Install Python 3.10 or higher
-
Create a Python virtual environment.
-
Inside that virtual environment, install the requirements:
python -m pip install -r requirements.txt
It's best to use a GPT-4 model for performing the evaluation, even if your chat app uses GPT-3.5 or another model. You can either use an Azure OpenAI instance or an openai.com instance.
To use a new Azure OpenAI instance, you'll need to create a new instance and deploy the app to it.
We've made that easy to deploy with the azd CLI tool.
-
Install the Azure Developer CLI
-
Run
azd auth loginto log in to your Azure account -
Run
azd upto deploy a new GPT-4 instance -
Create a
.envfile based on the provisioned resources by running one of the following commands.Bash:
azd env get-values > .envPowerShell:
$output = azd env get-values; Add-Content -Path .env -Value $output;
If you already have an Azure OpenAI instance, you can use that instead of creating a new one.
-
Create
.envfile by copying.env.sample -
Fill in the values for your instance:
AZURE_OPENAI_EVAL_DEPLOYMENT="<deployment-name>" AZURE_OPENAI_SERVICE="<service-name>"
-
The scripts default to keyless access (via
AzureDefaultCredential), but you can optionally use a key by settingAZURE_OPENAI_KEYin.env.
If you have an openai.com instance, you can use that instead of an Azure OpenAI instance.
-
Create
.envfile by copying.env.sample -
Change
OPENAI_HOSTto "openai" and fill in the key for for your OpenAI account. If you do not have an organization, you can leave that blank.OPENAI_HOST="openai" OPENAICOM_KEY="" OPENAICOM_ORGANIZATION=""
In order to evaluate new answers, they must be compared to "ground truth" answers: the ideal answer for a particular question. See example_input/qa.jsonl for an example of the format.
We recommend at least 200 QA pairs if possible.
There are a few ways to get this data:
- Manually curate a set of questions and answers that you consider to be ideal. This is the most accurate, but also the most time-consuming. Make sure your answers include citations in the expected format. This approach requires domain expertise in the data.
- Use the generator script to generate a set of questions and answers. This is the fastest, but may also be the least accurate. See below for details on how to run the generator script.
- Use the generator script to generate a set of questions and answers, and then manually curate them, rewriting any answers that are subpar and adding missing citations. This is a good middle ground, and is what we recommend.
Additional tips for ground truth data generation
- Generate more QA pairs than you need, then prune them down manually based on quality and overlap. Remove low quality answers, and remove questions that are too similar to other questions.
- Be aware of the knowledge distribution in the document set, so you effectively sample questions across the knowledge space.
- Once your chat application is live, continually sample live user questions (within accordance to your privacy policy) to make sure you're representing the sorts of questions that users are asking.
This repo includes a script for generating questions and answers from documents stored in Azure AI Search.
Important
The generator script can only generate English Q/A pairs right now, due to limitations in the azure-ai-generative SDK.
-
Create
.envfile by copying.env.sample -
Fill in the values for your Azure AI Search instance:
AZURE_SEARCH_SERVICE="<service-name>" AZURE_SEARCH_INDEX="<index-name>" AZURE_SEARCH_KEY=""
The key may not be necessary if it's configured for keyless access from your account. If providing a key, it's best to provide a query key since the script only requires that level of access.
-
Run the generator script:
python -m scripts generate --output=example_input/qa.jsonl --numquestions=200 --persource=5
That script will generate 200 questions and answers, and store them in
example_input/qa.jsonl. We've already provided an example based off the sample documents for this app.To further customize the generator beyond the
numquestionsandpersourceparameters, modifyscripts/generate.py.
We provide a script that loads in the current azd environment's variables, installs the requirements for the evaluation, and runs the evaluation against the local app. Run it like this:
python -m scripts evaluate --config=example_config.jsonThe config.json should contain these fields as a minimum:
{
"testdata_path": "example_input/qa.jsonl",
"target_url": "http://localhost:50505/chat",
"requested_metrics": ["groundedness", "relevance", "coherence", "latency", "answer_length"],
"results_dir": "example_results/experiment<TIMESTAMP>"
}If you're running this evaluator in a container and your app is running in a container on the same system, use a URL like this for the target_url:
"target_url": "http://host.docker.internal:50505/chat"
To run against a deployed endpoint, change the target_url to the chat endpoint of the deployed app:
"target_url": "https://app-backend-j25rgqsibtmlo.azurewebsites.net/chat"
It's common to run the evaluation on a subset of the questions, to get a quick sense of how the changes are affecting the answers. To do this, use the --numquestions parameter:
python -m scripts evaluate --config=example_config.json --numquestions=2The evaluate command will use the metrics specified in the requested_metrics field of the config JSON.
Some of those metrics are built-in to the evaluation SDK, and the rest are custom metrics that we've added.
These metrics are calculated by sending a call to the GPT model, asking it to provide a 1-5 rating, and storing that rating.
Important
The built-in metrics are only intended for use on evaluating English language answers, since they use English-language prompts internally. For non-English languages, you should use the custom prompt metrics instead.
gpt_coherencemeasures how well the language model can produce output that flows smoothly, reads naturally, and resembles human-like language.gpt_relevanceassesses the ability of answers to capture the key points of the context.gpt_groundednessassesses the correspondence between claims in an AI-generated answer and the source context, making sure that these claims are substantiated by the context.gpt_similaritymeasures the similarity between a source data (ground truth) sentence and the generated response by an AI model.gpt_fluencymeasures the grammatical proficiency of a generative AI's predicted answer.f1_scoreMeasures the ratio of the number of shared words between the model generation and the ground truth answers.
The following metrics are implemented very similar to the built-in metrics, but use a locally stored prompt. They're a great fit if you find that the built-in metrics are not working well for you or if you need to translate the prompt to another language.
mycoherence: Measures how well the language model can produce output that flows smoothly, reads naturally, and resembles human-like language. Based onscripts/evaluate_metrics/prompts/coherence.prompty.myrelevance: Assesses the ability of answers to capture the key points of the context. Based onscripts/evaluate_metrics/prompts/relevance.prompty.mygroundedness: Assesses the correspondence between claims in an AI-generated answer and the source context, making sure that these claims are substantiated by the context. Based onscripts/evaluate_metrics/prompts/groundedness.prompty.
These metrics are calculated with some local code based on the results of the chat app, and do not require a call to the GPT model.
latency: The time it takes for the chat app to generate an answer, in seconds.length: The length of the generated answer, in characters.has_citation: Whether the answer contains a correctly formatted citation to a source document, assuming citations are in square brackets.citation_match: Whether the answer contains at least all of the citations that were in the ground truth answer.
This repo assumes that your chat app is following the Chat App Protocol, which means that all POST requests look like this:
{"messages": [{"content": "<Actual user question goes here>", "role": "user"}],
"stream": False,
"context": {...},
}Any additional app parameters would be specified in the context of that JSON, such as temperature, search settings, prompt overrides, etc. To specify those parameters, add a target_parameters key to your config JSON. For example:
"target_parameters": {
"overrides": {
"semantic_ranker": false,
"prompt_template": "<READFILE>example_input/prompt_refined.txt"
}
}The overrides key is the same as the overrides key in the context of the POST request.
As a convenience, you can use the <READFILE> prefix to read in a file and use its contents as the value for the parameter.
That way, you can store potential (long) prompts separately from the config JSON file.
The results of each evaluation are stored in a results folder (defaulting to example_results).
Inside each run's folder, you'll find:
eval_results.jsonl: Each question and answer, along with the GPT metrics for each QA pair.parameters.json: The parameters used for the run, like the overrides.summary.json: The overall results, like the average GPT metrics.config.json: The original config used for the run. This is useful for reproducing the run.
To make it easier to view and compare results across runs, we've built a few tools,
located inside the review-tools folder.
To view a summary across all the runs, use the summary command with the path to the results folder:
python -m review_tools summary example_resultsThis will display an interactive table with the results for each run, like this:
To see the parameters used for a particular run, select the folder name. A modal will appear with the parameters, including any prompt override.
To compare the answers generated for each question across 2 runs, use the compare command with 2 paths:
python -m review_tools diff example_results/baseline_1 example_results/baseline_2This will display each question, one at a time, with the two generated answers in scrollable panes, and the GPT metrics below each answer.
]
Use the buttons at the bottom to navigate to the next question or quit the tool.
You can also filter to only show questions where the value changed for a particular metric, like this:
python -m review_tools diff example_results/baseline_1 example_results/baseline_2 --changed=has_citationThe evaluation flow described above focused on evaluating a model’s answers for a set of questions that could be answered by the data. But what about all those questions that can’t be answered by the data? Does your model know how to say “I don’t know?” The GPT models are trained to try and be helpful, so their tendency is to always give some sort of answer, especially for answers that were in their training data. If you want to ensure your app can say “I don’t know” when it should, you need to evaluate it on a different set of questions with a different metric.
For this evaluation, our ground truth data needs to be a set of question whose answer should provoke an "I don’t know" response from the data. There are several categories of such questions:
- Unknowable: Questions that are related to the sources but not actually in them (and not public knowledge).
- Uncitable: Questions whose answers are well known to the LLM from its training data, but are not in the sources. There are two flavors of these:
- Related: Similar topics to sources, so LLM will be particularly tempted to think the sources know.
- Unrelated: Completely unrelated to sources, so LLM shouldn’t be as tempted to think the sources know.
- Nonsensical: Questions that are non-questions, that a human would scratch their head at and ask for clarification.
You can write these questions manually, but it’s also possible to generate them using a generator script in this repo, assuming you already have ground truth data with answerable questions.
python -m scripts generate-dontknows --input=example_input/qa.jsonl --output=example_input/qa_dontknows.jsonl --numquestions=45That script sends the current questions to the configured GPT-4 model along with prompts to generate questions of each kind.
When it’s done, you should review and curate the resulting ground truth data. Pay special attention to the "unknowable" questions at the top of the file, since you may decide that some of those are actually knowable, and you may want to reword or rewrite entirely.
This repo contains a custom GPT metric called "dontknowness" that rates answers from 1-5, where 1 is "answered the question completely with no certainty" and 5 is "said it didn't know and attempted no answer". The goal is for all answers to be rated 4 or 5.
Here's an example configuration JSON that requests that metric, referencing the new ground truth data and a new output folder:
{
"testdata_path": "example_input/qa_dontknows.jsonl",
"results_dir": "example_results_dontknows/baseline",
"requested_metrics": ["dontknowness", "answer_length", "latency", "has_citation"],
"target_url": "http://localhost:50505/chat",
"target_parameters": {
}
}We recommend a separate output folder, as you'll likely want to make multiple runs and easily compare between those runs using the review tools.
Run the evaluation like this:
python -m scripts evaluate --config=dontknows.config.jsonThe results will be stored in the results_dir folder, and can be reviewed using the review tools.
If the app is not saying "I don't know" enough, you can use the diff tool to compare the answers for the "dontknows" questions across runs, and see if the answers are improving. Changes you can try:
- Adjust the prompt to encourage the model to say "I don't know" more often. Remove anything in the prompt that might be distracting or overly encouraging it to answer.
- Try using GPT-4 instead of GPT-3.5. The results will be slower (see the latency column) but it may be more likely to say "I don't know" when it should.
- Adjust the temperature of the model used by your app.
- Add an additional LLM step in your app after generating the answer, to have the LLM rate its own confidence that the answer is found in the sources. If the confidence is low, the app should say "I don't know".
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
![microsoft-github-operations[bot] avatar](https://avatars.githubusercontent.com/in/41902?v=4 "microsoft-github-operations[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent