This repository contains the code that Defog uses for the evaluation of generated SQL. It's based off the schema from the Spider, but with a new set of hand-selected questions and queries grouped by query category. For an in-depth look into our process of creating this evaluation approach, see this.
Our testing procedure comprises the following steps. For each question/query pair:
- We generate a SQL query (possibly from an LLM).
- We run both the "gold" query and the generated query on their respective database to obtain 2 dataframes with the results.
- We compare the 2 dataframes using an "exact" and a "subset" match. TODO add link to blogpost.
- We log these alongside other metrics of interest (e.g. tokens used, latency) and aggregate the results for reporting.
This is a comprehensive set of instructions that assumes basic familiarity with the command line, Docker, running SQL queries on a database, and common Python data manipulation libraries (e.g. pandas).
Firstly, clone the repository where we store our database data and schema. Install all Python libraries listed in the requirements.txt
file. You would also need to download a spacy model if you're using the NER heuristic for our metadata-pruning method (set by any values of the c
parameter that is more than 0, more below). Finally, install the library.
git clone https://github.com/defog-ai/defog-data.git
cd defog-data
pip install -r requirements.txt
python -m spacy download en_core_web_sm
pip install -e .
Next, you would need to set up the databases that the queries are executed on. We use Postgres here, since it is the most common OSS database with the widest distribution and usage in production. In addition, we would recommend using Docker to do this, as it is the easiest way to get started. You can install Docker here.
Once you have Docker installed, you can create the Docker container and start the Postgres database using the following commands. We recommend mounting a volume on data/postgres
to persist the data, as well as data/export
to make it easier to import the data. To create the container, run:
mkdir data/postgres data/export
docker create --name postgres-sql-eval -e POSTGRES_PASSWORD=postgres -p 5432:5432 -v $(pwd)/data/postgres:/var/lib/postgresql/data -v $(pwd)/data/export:/export postgres:16-alpine
To start the container, run:
docker start postgres-sql-eval
If you want to reset the Postgres server instance's state (e.g. memory leaks from transient connections), you can turn it off (and start it back up after):
docker stop postgres-sql-eval
# see that the container is still there:
docker container list -a
Some notes:
- You would need to stop other Postgres instances listening on port 5432 before running the above command.
- You only need to run the
docker create ...
once to create the image, and then subsequently onlydocker start/stop postgres-sql-eval
. - The data is persisted in
data/postgres
, so turning it off isn't critical. On the other hand, if you delete thedata/postgres
folder, then all is lost T.T - While we will use Docker for deploying Postgres and the initialization, you are free to modify the scripts/instructions to work with your local installation.
The data for importing is in the defog-data
repository which we cloned earlier. Each folder contains the metadata and data corresponding to a single database (e.g. academic
contains all the data required to reload the 'academic' database). We assume that you have a psql
client installed locally. We will create a new database in our postgres instance for each of the 7 SQL databases with the following commands:
# set the following environment variables
cd defog-data # if you're not already in the defog-data directory
export DBPASSWORD="postgres"
export DBUSER="postgres"
export DBHOST="localhost"
export DBPORT=5432
./setup.sh
Should you wish to import the data into Snowflake, the setup instructions are also in the defog-data
repository. After installing the Snowflake CLI, configure your credentials as per the docs and set them as environment variables like below, then run the setup command.
export SFDBPASSWORD="your_password"
export SFDBUSER="your_username"
export SFDBACCOUNT="your_account"
export SFDBWAREHOUSE="your_warehouse"
./setup_snowflake.sh
Note that during evaluation you'll have to use the _snowflake
question files in /data
. The queries been modified to be valid on Snowflake databases.
The setup instructions for these database management systems are found in the defog-data
repository. Configure your credentials accordingly, set up your environment variables, then translate and import the eval databases with the command:
python translate_ddl_dialect.py
During evaluation, you'll have to set the right --db_type
flag and use the corresponding _{dialect}
question files in /data
.
If you have a private dataset that you do not want to make publicly available but would still like to repurpose the code here for evaluations, you can do so by following the steps below.
- Begin by creating a separate git repository for your private data, that has a
setup.py
file, similar to defog-data. - Create the metadata and data files, and import them into your database. This is to allow our evaluation framework to run the generated queries with some actual data. You can refer to
defog-data
's metadata objects for the schema, and setup.sh as an example on how import the data into your database. We do not prescribe any specific folder structure, and leave it to you to decide how you want to organize your data, so long as you can import it into your database easily. - To use our metadata pruning utilities, you would need to have the following defined:
- A way to load your embeddings. In our case, we call a function load_embeddings from
defog-data
's supplementary module to load a dictionary of database name to a tuple of the 2D embedding matrix (num examples x embedding dimension) and the associated text metadata for each row/example. If you would like to see how we generate this tuple, you may refer to generate_embeddings in thedefog-data
repository. - A way to load columns associated with various named entities. In our case, we call a dictionary columns_ner of database name to a nested dictionary that maps each named entity type to a list of column metadata strings that are associated with that named entity type. You can refer to the raw data for an example of how we generate this dictionary.
- A way to define joinable columns between tables. In our case, we call a dictionary columns_join of database name to a nested dictionary of table tuples to column name tuples. You can refer to the raw data for an example of how we generate this dictionary.
- A way to load your embeddings. In our case, we call a function load_embeddings from
Once all of the 3 above steps have completed, you would need to
- Install your data library as a dependency, by running
pip install -e .
(-e to automatically incorporate edits without reinstalling) - Replace the associated function calls and variables in prune_metadata_str with your own imported functions and variables. Note that you might not name your package/module
defog_data_private.supplementary
, so do modify accordingly.
Some things to take note of:
- If you do not populate your database with data (ie only create the tables without inserting data), you would return empty dataframes most of the time (regardless of whether the query generated was what you want), and it would result in results matching all the time and generate a lot of false positives. Hence, you might want to consider populating your database with some meaningful data that would return different results if the queries should be different from what you want.
- If testing out on your private data, you would also need to change the questions file to point to your own questions file (tailored to your database schema).
To test your own query generator with our framework, you would need to extend Query Generator and implement the generate_query method to return the query of interest. We create a new class for each question/query pair to isolate each pair's runtime state against the others when running concurrently. You can also reference OpenAIQueryGenerator which implements Query Generator
and uses a simple prompt to send a message over to OpenAI's API. Feel free to extend it for your own use.
If there are functions that are generally useful for all query generators, they can be placed in the utils
folder. If you need to incorporate specific verbose templates (e.g. for prompt testing), you can store them in the prompts
folder, and later import them. Being able to version control the prompts in a central place has been a productivity win for our team.
Having implemented the query generator, the next piece of abstraction would be the runner. The runner calls the query generator, and is responsible for handling the configuration of work (e.g. parallelization / batching / model selected etc.) to the query generator for each question/query pair.
We have provided a few common runners: eval/openai_runner.py
for calling OpenAI's API (with parallelization support), eval/anthropic_runner
for calling Anthropic's API, eval/hf_runner.py
for calling a local Hugging Face model and finally, eval/api_runner.py
makes it possible to use a custom API for evaluation.
When testing your own query generator with an existing runner, you can replace the qg_class
in the runner's code with your own query generator class.
Remember to have your API key (OPENAI_API_KEY
or ANTHROPIC_API_KEY
) set as an environment variable before running the test if you plan to call the OpenAI or Anthropic/other LLM API's accordingly.
To test it out with just 10 questions (instead of all 200), parallelized across 5 :
python main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" "data/instruct_basic_postgres.csv" "data/instruct_advanced_postgres.csv" \
-o results/openai_classic.csv results/openai_basic.csv results/openai_advanced.csv \
-g oa \
-f prompts/prompt_openai.json \
-m gpt-4-turbo \
-p 5 \
-c 0
If testing with the latest o1-*
models (which do not support system prompts), you should use a different prompt file, reduce parallel requests and increase the timeout:
python main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" "data/instruct_basic_postgres.csv" "data/instruct_advanced_postgres.csv" \
-o results/openai_o1mini_classic.csv results/openai_o1mini_basic.csv results/openai_o1mini_advanced.csv \
-g oa \
-f prompts/prompt_openai_o1.json \
-m o1-mini \
-p 1 \
-t 120 \
-c 0
To test out the full suite of questions for claude-3:
python main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" "data/instruct_basic_postgres.csv" "data/instruct_advanced_postgres.csv" \
-o results/claude3_classic.csv results/claude3_basic.csv results/claude3_advanced.csv \
-g anthropic \
-f prompts/prompt_anthropic.md \
-m claude-3-opus-20240229 \
-p 5 \
-c 0
To test it out with our fine-tuned sql model with just 10 questions (instead of all 200):
# use the -W option to ignore warnings about sequential use of transformers pipeline
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" "data/instruct_basic_postgres.csv" "data/instruct_advanced_postgres.csv" \
-o results/hf_classic.csv results/hf_basic.csv results/hf_advanced.csv \
-g hf \
-f prompts/prompt.md \
-m defog/llama-3-sqlcoder-8b \
-c 0
We also support loading a peft adapter here as well via the -a
flag. Note that the loading of the adapter with the model will take slightly longer than usual.
We also have a vllm runner which uses the vLLM engine to run the inference altogether as a single batch. It is much faster to do so especially when num_beams
> 1. You would have to pass in a single set of merged model weights, path to LoRA adapters if applicable, and the model architecture needs to be supported by vLLM. Here's a sample command:
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" "data/instruct_basic_postgres.csv" "data/instruct_advanced_postgres.csv" \
-o results/vllm_classic.csv results/vllm_basic.csv results/vllm_advanced.csv \
-g vllm \
-f "prompts/prompt.md" \
-m defog/llama-3-sqlcoder-8b \
-a path/to_adapter \
-c 0
Optionally, if you're running evals on a model that is quantized with AWQ, add the -qz
or --quantized
parameter. Only applicable for the vllm runner.
If running with different settings, you can setup an api server to avoid reloading for each test setting and then run the tests subsequently. We enable setting up 2 types of api servers, namely the vllm api server, as well as the TGI server.
We also provide our custom modification of the vllm api server, which only returns the generated output.
# to set up a vllm server
python -m vllm.entrypoints.api_server \
--model defog/defog-llama-3-sqlcoder-8b \
--tensor-parallel-size 4 \
--dtype float16
# to set up a vllm server that supports LoRA adapters
python -m vllm.entrypoints.api_server \
--model defog/llama-3-sqlcoder-8b \
--tensor-parallel-size 1 \
--dtype float16 \
--max-model-len 4096 \
--enable-lora \
--max-lora-rank 64
# to use our modified api server
python utils/api_server.py \
--model defog/llama-3-sqlcoder-8b \
--tensor-parallel-size 4 \
--dtype float16 \
--max-model-len 4096 \
--enable-lora \
--max-lora-rank 64
# to run sql-eval using the api runner - depending on how much your GPUs can take, can increase p and b to higher values
python main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o results/api.csv \
-g api \
-b 1 \
-f prompts/prompt.md \
--api_url "http://localhost:8000/generate" \
--api_type "vllm" \
-a path/to_adapter_if_applicable \
-p 8
You may consult the TGI documentation for more information on how to set up a TGI server. Here's a sample command to set up a TGI server using a preset docker image and run the evaluation using the API runner. Note that you would want to change the number of shards and the model id accordingly, depending on how many gpu's you have available and your model of choice.
# to set up a tgi server
model="defog/llama-3-sqlcoder-8b"
docker run --gpus all \
--shm-size 1g \
-p 8000:80 \
-v /models:/models ghcr.io/huggingface/text-generation-inference:2.0 \
--model-id "${model}" \
--max-best-of 4 \
--max-input-tokens 3072 \
--sharded true \
--num-shard 4 \
--hostname 0.0.0.0 \
--port 80
# to run sql-eval using the api runner - depending on how much your GPUs can take, can increase p and b to higher values. Note that cuda graphs in tgi is optimized for batch sizes that are powers of 2 by default.
python main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o results/api.csv \
-g api \
-b 1 \
-f prompts/prompt.md \
--api_url "http://localhost:8000/generate" \
--api_type "vllm" \
-p 8
If you'd like to test out a few prompts in a single run (to save the few minutes spent loading the model into GPU at the start of each run), you can specify a list of prompt files in --prompt_file
(e.g. -f prompts/prompt-1.md prompts/prompt-2.md prompts/prompt-3.md
), as well as a corresponding list of output files in --output_file
(e.g. -o results/results-1.csv results/results-2.csv results/results-3.csv
). The number of prompts and output files must be the same. Here's a sample command:
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o results/results_1.csv results/results_2.csv \
-g vllm \
-f prompts/prompt_1.md prompts/prompt_2.md \
-m defog/sqlcoder2
While you can do the same for the other runners, the time savings are most significant when loading a large model locally, vs calling an always-on API.
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o results/llama3_70b.csv \
-g bedrock \
-f prompts/prompt.md \
-m meta.llama3-70b-instruct-v1:0
To run the eval using Llama CPP, you can use the following code. Before running this, you must install llama-cpp-python
with the following (on Apple Silicon)
CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python
Note that llama-cpp-python library does not currently have beam search, and hence will have lower quality results.
python -W ignore main.py \
-q "data/questions_gen_postgres.csv" \
-db postgres \
-o "results/llama_cpp.csv" \
-g llama_cpp \
-f "prompts/prompt.md" \
-m path/to/model.gguf
To run the eval using MLX, you can use the following code. Before running this, you must install mlx-lm
package with pip install mlx-lm
Note that MLX does not currently have beam search, and hence will have lower quality results.
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o "results/mlx_llama-3-sqlcoder-8b.csv" \
-g mlx \
-f "prompts/prompt.md" \
-m mlx-community/defog-llama-3-sqlcoder-8b
Before running this, you must create an account with Google AI and set your credentials with export GOOGLE_APPLICATION_CREDENTIALS=</path/to/service_account.json>
. Then, install these packages with pip install vertexai google-cloud-aiplatform
.
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o "results/gemini_pro.csv" \
-g gemini \
-f "prompts/prompt_gemini.md" \
-m gemini-pro \
-p 1 \
-n 5
Before running this, you must create an account with Mistral and obtain an API key and store it with export MISTRAL_API_KEY=<your_api_key>
. Then, install mistralai
with pip install mistralai
.
python -W ignore main.py \
-db postgres \
-q "data/questions_gen_postgres.csv" \
-o "results/results.csv" \
-g mistral \
-f "prompts/prompt_mistral.md" \
-m mistral-medium \
-p 5 \
-n 10
Before running this, you would need to export the following environment variables for the boto3 client to work:
AWS_ACCESS_KEY_ID
AWS_SECRET_ACCESS_KEY
AWS_DEFAULT_REGION
python3 main.py \
-db postgres \
-q data/instruct_basic_postgres.csv data/instruct_advanced_postgres.csv data/questions_gen_postgres.csv \
-o results/bedrock_llama_70b_basic.csv results/bedrock_llama_70b_advanced.csv results/bedrock_llama_70b_v1.csv \
-g bedrock \
-f prompts/prompt_cot_postgres.md \
-m meta.llama3-70b-instruct-v1:0 \
-c 0 \
-p 10
Before running this, you must create an account with Together.ai and obtain an API key and store it with export TOGETHER_API_KEY=<your_api_key>
. Then, install together
with pip install together
. You can then run the following command:
python3 main.py \
-db postgres \
-q data/instruct_basic_postgres.csv data/instruct_advanced_postgres.csv data/questions_gen_postgres.csv \
-o results/together_llama_70b_basic.csv results/together_llama_70b_advanced.csv results/together_llama_70b_v1.csv \
-g together \
-f prompts/prompt_together.json \
-m "meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo" \
-c 0 \
-p 10
You can use the following flags in the command line to change the configurations of your evaluation runs.
CLI Flags | Description |
---|---|
-q, --questions_file | CSV file that contains the test questions and true queries. If this is not set, it will default to the relevant questions_gen_<db_type>.csv file. It may be helpful to always end your questions*file name with *<db_type>.csv to ensure compatibility between the queries and selected db_type. |
-n, --num_questions | Use this to limit the total number of questions you want to test. |
-db, --db_type | Database type to run your queries on. Currently supported types are postgres and snowflake . |
-d, --use_private_data | Use this to read from your own private data library. |
-dp, --decimal_points | Use this to specify the number of decimal points a result should be rounded to. This is None by default |
CLI Flags | Description |
---|---|
-g, --model_type | Model type used. Make sure this matches the model used. Currently defined options in main.py are oa for OpenAI models, anthropic for Anthropic models, hf for Hugging Face models, vllm for a vllm runner, api for API endpoints, llama_cpp for llama cpp, mlx for mlx, bedrock for AWS bedrock API, together for together.ai's API |
-m, --model | Model that will be tested and used to generate the queries. Some options for OpenAI models are chat models gpt-3.5-turbo-0613 and gpt-4-0613 . Options for Anthropic include the latest claude-3 family of models (e.g. claude-3-opus-20240229 ). For Hugging Face, and VLLM models, simply use the path of your chosen model (e.g. defog/sqlcoder ). |
-a, --adapter | Path to the relevant adapter model you're using. Only available for the hf_runner . |
--api_url | The URL of the custom API you want to send the prompt to. Only used when model_type is api . |
-qz, --quantized | Indicate whether the model is an AWQ quantized model. Only available for vllm_runner . |
CLI Flags | Description | |
---|---|---|
-f, --prompt_file | Markdown file with the prompt used for query generation. You can pass in a list of prompts to test sequentially without reloading the script. | |
-b, --num_beams | Indicates the number of beams you want to use for beam search at inference. Only available for hf_runner , vllm_runner , and api_runner . |
|
-c, --num_columns | Number of columns, default 20. To not prune the columns, set it to 0. | |
-s, --shuffle_metadata | Shuffle metadata, default False. This shuffles the order of the tables within the schema and the order of the columns within each table but does not shift columns between tables (to preserve the structure of the database). | |
-k, --k_shot | Used when you want to include k-shot examples in your prompt. Make sure that the column 'k_shot_prompt' exists in your questions_file. | |
--cot_table_alias | (Experimental) Used when you want to include chain-of-thought instructions before the actual sql generation. Allowed values are instruct . If using instruct , make sure that the placeholder '{cot_instructions}' exists in your prompt file. instruct will get your model generate the chain-of-thought table aliases. |
CLI Flags | Description |
---|---|
-o, --output_file | Output CSV file that will store your results. You need to pass the same number of output file paths as the number of prompt files. |
-p, --parallel_threads | No. of parallel workers available for generating and processing queries |
-t, --timeout_gen | No. of seconds before timeout occurs for query generation. The default is 30.0s. |
-u, --timeout_exec | No. of seconds before timeout occurs for query execution on the database. The default is 10.0s. |
-v, --verbose | Prints details in command line. |
--upload_url | (optional) the URL that you want to report the results to. The server that serves this URL must have functionality that is similar to the sample server in utils/webserver.py . |
--run_name | (optional) the name of this run for logging purposes |
To better understand your query generator's performance, you can explore the results generated and aggregated for the various metrics that you care about.
If you would like to start a google cloud function to receive the results, you can use the --upload_url
flag to specify the URL that you want to report the results to. Before running the evaluation code with this flag, you would need to create a server that serves at the provided URL. We have provided 2 sample cloud function endpoints for writing either to bigquery or postgres, in the results_fn_bigquery
and results_fn_postgres
folders. You may also implement your own server to take in similar arguments. Before deploying either cloud functions, you would need to set up the environment variables by making a copy of .env.yaml.template and renaming it to .env.yaml, and then filling in the relevant fields. For the bigquery cloud function, you would also need to put your service account's key.json file in the same folder, and put the file name in the CREDENTIALS_PATH
field in the .env.yaml file.
After doing so, you can deploy the google cloud function:
# for uploading to bigquery
gcloud functions deploy results_bigquery \
--source results_fn_bigquery \
--entry-point bigquery \
--env-vars-file results_fn_bigquery/.env.yaml \
--runtime python311 \
--memory 512MB \
--trigger-http \
--allow-unauthenticated \
--gen2
# for uploading to postgres
gcloud functions deploy results_postgres \
--source results_fn_postgres \
--entry-point postgres \
--env-vars-file results_fn_postgres/.env.yaml \
--runtime python311 \
--memory 512MB \
--trigger-http \
--allow-unauthenticated \
--gen2
The cloud function's name is whatever comes after gcloud functions deploy
(in this case, results_bigquery
), and you can use it to check the logs of the function by running gcloud functions logs read results_bigquery
.
You can then run the evaluation code with the --upload_url
flag to report the results to the cloud function. The cloud function will then write the results to the relevant database.
python main.py \
-db postgres \
-o results/test.csv \
-g oa \
-f prompts/prompt_openai.json \
-m gpt-3.5-turbo-0613 \
-n 1 \
--upload_url <your cloud function url>
If you would like to always report your results to an upload_url, even if it's not explicitly provided, you can set it in your environment variables as SQL_EVAL_UPLOAD_URL
If you'd like to modify the functions and test it out locally, you can run these sample commands to deploy the function locally and then trigger the openai runner:
functions-framework --target bigquery --source results_fn_bigquery --debug
python main.py \
-db postgres \
-o results/test.csv \
-g oa \
-f prompts/prompt_openai.json \
-m gpt-3.5-turbo-0613 \
-n 1 \
--upload_url http://127.0.0.1:8080/
We welcome contributions to our project, specifically:
- Dataset
- Adding new database schema/data
- Framework code
- New query generators/runners (in the query_generators and eval folders respectively)
- Improving existing generators/runners (e.g. adding new metrics)
Please see CONTRIBUTING.md for more information.