Awesome

SQL Generation Evaluation

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.

Introduction

Our testing procedure comprises the following steps. For each question/query pair:

1. We generate a SQL query (possibly from an LLM).
2. We run both the "gold" query and the generated query on their respective database to obtain 2 dataframes with the results.
3. We compare the 2 dataframes using an "exact" and a "subset" match. TODO add link to blogpost.
4. We log these alongside other metrics of interest (e.g. tokens used, latency) and aggregate the results for reporting.

Getting Started

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).

Install Dependencies

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 .

Start Postgres Instance

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 only docker 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 the data/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.

Import Data into Postgres

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

Import Data into Snowflake

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.

Import Data into BigQuery, MySQL, SQLite, SQL Server

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.

Using Private Data (Optional)

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 the defog-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.

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).

Query Generator

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.

Runner

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.

Running the Test

OpenAI

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

Anthropic

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

Hugging Face

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.

vLLM

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.

Running with an API Server

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.

VLLM API Server

# 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

TGI API Server

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

Multiple Prompts

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.

Bedrock

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

Llama CPP

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

MLX

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

Gemini

Before running this, you need to set your credentials with export GEMINI_API_KEY=<your_api_key>. Then, install these packages with pip install google-generative-ai.

python main.py \
  -db postgres \
  -q "data/questions_gen_postgres.csv" "data/instruct_basic_postgres.csv" "data/instruct_advanced_postgres.csv" \
  -o "results/gemini_flash_basic.csv" "results/gemini_flash_basic.csv" "results/gemini_flash_advanced.csv" \
  -g gemini \
  -f "prompts/prompt_gemini.md" "prompts/prompt_gemini.md" "prompts/prompt_gemini.md" \
  -m gemini-2.0-flash-exp \
  -p 10

Mistral

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

Bedrock

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

Together

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

CLI Flags

You can use the following flags in the command line to change the configurations of your evaluation runs.

Data-related parameters

Model-related parameters

Inference-technique-related parameters

Execution-related parameters

Checking the Results

To better understand your query generator's performance, you can explore the results generated and aggregated for the various metrics that you care about.

Upload URL

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

Testing the function locally

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/

Misc

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.