blendsql 0.0.31

Query language for blending SQL logic and LLM reasoning across multi-modal data. [Findings of ACL 2024]

SQL 🤝 LLMs

Check out our online documentation for a more comprehensive overview.

Results from the paper are available here

pip install blendsql --prerelease=allow

✨ News

• (10/26/24) New tutorial! blendsql-by-example.ipynb
• (10/18/24) Concurrent async requests in 0.0.29! OpenAI and Anthropic LLMMap calls are speedy now.
  • Customize max concurrent async calls via blendsql.config.set_async_limit(10)
• (10/15/24) As of version 0.0.27, there is a new pattern for defining + retrieving few-shot prompts; check out Few-Shot Prompting in the README for more info
• (10/15/24) Check out Some Cool Things by Example for some recent language updates!

Features

• Supports many DBMS 💾
  • SQLite, PostgreSQL, DuckDB, Pandas (aka duckdb in a trenchcoat)
• Supports many models ✨
  • Transformers, OpenAI, Anthropic, Ollama
• Easily extendable to multi-modal usecases 🖼️
• Write your normal queries - smart parsing optimizes what is passed to external functions 🧠
  • Traverses abstract syntax tree with sqlglot to minimize LLM function calls 🌳
• Constrained decoding with guidance 🚀
  • When using local models, we only generate syntactically valid outputs according to query syntax + database contents
• LLM function caching, built on diskcache 🔑

BlendSQL is a superset of SQLite for problem decomposition and hybrid question-answering with LLMs.

As a result, we can Blend together...

• 🥤 ...operations over heterogeneous data sources (e.g. tables, text, images)
• 🥤 ...the structured & interpretable reasoning of SQL with the generalizable reasoning of LLMs

Now, the user is given the control to oversee all calls (LLM + SQL) within a unified query language.

For example, imagine we have the following table titled parks, containing info on national parks in the United States.

We can use BlendSQL to build a travel planning LLM chatbot to help us navigate the options below.

Name	Image	Location	Area	Recreation Visitors (2022)	Description
Death Valley		California, Nevada	3,408,395.63 acres (13,793.3 km2)	1,128,862	Death Valley is the hottest, lowest, and driest place in the United States, with daytime temperatures that have exceeded 130 °F (54 °C).
Everglades		Alaska	7,523,897.45 acres (30,448.1 km2)	9,457	The country's northernmost park protects an expanse of pure wilderness in Alaska's Brooks Range and has no park facilities.
New River Gorge		West Virgina	7,021 acres (28.4 km2)	1,593,523	The New River Gorge is the deepest river gorge east of the Mississippi River.
Katmai		Alaska	3,674,529.33 acres (14,870.3 km2)	33,908	This park on the Alaska Peninsula protects the Valley of Ten Thousand Smokes, an ash flow formed by the 1912 eruption of Novarupta.

BlendSQL allows us to ask the following questions by injecting "ingredients", which are callable functions denoted by double curly brackets ({{, }}).

Which parks don't have park facilities?

SELECT "Name", "Description" FROM parks
  WHERE {{
      LLMMap(
          'Does this location have park facilities?',
          context='parks::Description'
      )
  }} = FALSE

Name	Description
Everglades	The country's northernmost park protects an expanse of pure wilderness in Alaska's Brooks Range and has no park facilities.

---

What does the largest park in Alaska look like?

SELECT "Name",
{{ImageCaption('parks::Image')}} as "Image Description", 
{{
    LLMMap(
        question='Size in km2?',
        context='parks::Area'
    )
}} as "Size in km" FROM parks
WHERE "Location" = 'Alaska'
ORDER BY "Size in km" DESC LIMIT 1

Name	Image Description	Size in km
Everglades	A forest of tall trees with a sunset in the background.	30448.1

---

Which state is the park in that protects an ash flow?

SELECT "Location", "Name" AS "Park Protecting Ash Flow" FROM parks 
    WHERE "Name" = {{
      LLMQA(
        'Which park protects an ash flow?',
        context=(SELECT "Name", "Description" FROM parks),
        options="parks::Name"
      ) 
  }}

Location	Park Protecting Ash Flow
Alaska	Katmai

---

How many parks are located in more than 1 state?

SELECT COUNT(*) FROM parks
    WHERE {{LLMMap('How many states?', 'parks::Location')}} > 1

Count
1

---

Give me some info about the park in the state that Sarah Palin was governor of.

SELECT "Name", "Location", "Description" FROM parks
  WHERE Location = {{RAGQA('Which state was Sarah Palin governor of?')}}

Name	Location	Description
Everglades	Alaska	The country's northernmost park protects an expanse of pure wilderness in Alaska's Brooks Range and has no park facilities.
Katmai	Alaska	This park on the Alaska Peninsula protects the Valley of Ten Thousand Smokes, an ash flow formed by the 1912 eruption of Novarupta.

---

What's the difference in visitors for those parks with a superlative in their description vs. those without?

SELECT SUM(CAST(REPLACE("Recreation Visitors (2022)", ',', '') AS integer)) AS "Total Visitors", 
{{LLMMap('Contains a superlative?', 'parks::Description', options='t;f')}} AS "Description Contains Superlative",
GROUP_CONCAT(Name, ', ') AS "Park Names"
FROM parks
GROUP BY "Description Contains Superlative"

Total Visitors	Description Contains Superlative	Park Names
43365	0	Everglades, Katmai
2722385	1	Death Valley, New River Gorge

---

Now, we have an intermediate representation for our LLM to use that is explainable, debuggable, and very effective at hybrid question-answering tasks.

For in-depth descriptions of the above queries, check out our documentation.

Quickstart

import pandas as pd

import blendsql
from blendsql.ingredients import LLMMap, LLMQA, LLMJoin
from blendsql.db import Pandas
from blendsql.models import TransformersLLM, OpenaiLLM, AnthropicLLM

# Optionally set how many async calls to allow concurrently
# This depends on your OpenAI/Anthropic/etc. rate limits
blendsql.config.set_async_limit(10)

# Load model
model = OpenaiLLM("gpt-4o-mini") # requires .env file with `OPENAI_API_KEY`
# model = AnthropicLLM("claude-3-haiku-20240307") # requires .env file with `ANTHROPIC_API_KEY`
# model = TransformersLLM('Qwen/Qwen1.5-0.5B') # run with any local Transformers model

# Prepare our local database
db = Pandas(
  {
    "w": pd.DataFrame(
      (
        ['11 jun', 'western districts', 'bathurst', 'bathurst ground', '11-0'],
        ['12 jun', 'wallaroo & university nsq', 'sydney', 'cricket ground',
         '23-10'],
        ['5 jun', 'northern districts', 'newcastle', 'sports ground', '29-0']
      ),
      columns=['date', 'rival', 'city', 'venue', 'score']
    ),
    "documents": pd.DataFrame(
      (
        ['bathurst, new south wales',
         'bathurst /ˈbæθərst/ is a city in the central tablelands of new south wales , australia . it is about 200 kilometres ( 120 mi ) west-northwest of sydney and is the seat of the bathurst regional council .'],
        ['sydney',
         'sydney ( /ˈsɪdni/ ( listen ) sid-nee ) is the state capital of new south wales and the most populous city in australia and oceania . located on australia s east coast , the metropolis surrounds port jackson.'],
        ['newcastle, new south wales',
         'the newcastle ( /ˈnuːkɑːsəl/ new-kah-səl ) metropolitan area is the second most populated area in the australian state of new south wales and includes the newcastle and lake macquarie local government areas .']
      ),
      columns=['title', 'content']
    )
  }
)

# Write BlendSQL query
query = """
SELECT * FROM w
WHERE city = {{
    LLMQA(
        'Which city is located 120 miles west of Sydney?',
        (SELECT * FROM documents WHERE content LIKE '%sydney%'),
        options='w::city'
    )
}}
"""
smoothie = blendsql.blend(
  query=query,
  db=db,
  ingredients={LLMMap, LLMQA, LLMJoin},
  default_model=model,
  # Optional args below
  infer_gen_constraints=True,
  verbose=True
)
print(smoothie.df)
# ┌────────┬───────────────────┬──────────┬─────────────────┬─────────┐
# │ date   │ rival             │ city     │ venue           │ score   │
# ├────────┼───────────────────┼──────────┼─────────────────┼─────────┤
# │ 11 jun │ western districts │ bathurst │ bathurst ground │ 11-0    │
# └────────┴───────────────────┴──────────┴─────────────────┴─────────┘
print(smoothie.meta.prompts)
# [
#   {
#       'answer': 'bathurst',
#       'question': 'Which city is located 120 miles west of Sydney?',
#       'context': [
#           {'title': 'bathurst, new south wales', 'content': 'bathurst /ˈbæθərst/ is a city in the central tablelands of new south wales , australia . it is about...'},
#           {'title': 'sydney', 'content': 'sydney ( /ˈsɪdni/ ( listen ) sid-nee ) is the state capital of new south wales and the most populous city in...'}
#       ]
#    }
# ]

---

Citation

@article{glenn2024blendsql,
      title={BlendSQL: A Scalable Dialect for Unifying Hybrid Question Answering in Relational Algebra},
      author={Parker Glenn and Parag Pravin Dakle and Liang Wang and Preethi Raghavan},
      year={2024},
      eprint={2402.17882},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Few-Shot Prompting

For the LLM-based ingredients in BlendSQL, few-shot prompting can be vital. In LLMMap, LLMQA and LLMJoin, we provide an interface to pass custom few-shot examples and dynamically retrieve those top-k most relevant examples at runtime, given the current inference example.

LLMMap

• Default examples
• All possible fields

from blendsql import blend, LLMMap
from blendsql.ingredients.builtin import DEFAULT_MAP_FEW_SHOT

ingredients = {
    LLMMap.from_args(
        few_shot_examples=[
            *DEFAULT_MAP_FEW_SHOT,
            {
                "question": "Is this a sport?",
                "mapping": {
                    "Soccer": "t",
                    "Chair": "f",
                    "Banana": "f",
                    "Golf": "t"
                },
                # Below are optional
                "column_name": "Items",
                "table_name": "Table",
                "example_outputs": ["t", "f"],
                "options": ["t", "f"],
                "output_type": "boolean"
            }
        ],
        # Will fetch `k` most relevant few-shot examples using embedding-based retriever
        k=2,
        # How many inference values to pass to model at once
        batch_size=5,
    )
}
smoothie = blend(
    query=blendsql,
    db=db,
    ingredients=ingredients,
    default_model=model,
)

LLMQA

• Default examples
• All possible fields

from blendsql import blend, LLMQA
from blendsql.ingredients.builtin import DEFAULT_QA_FEW_SHOT

ingredients = {
    LLMQA.from_args(
        few_shot_examples=[
            *DEFAULT_QA_FEW_SHOT,
            {
                "question": "Which weighs the most?",
                "context": {
                    {
                        "Animal": ["Dog", "Gorilla", "Hamster"],
                        "Weight": ["20 pounds", "350 lbs", "100 grams"]
                    }
                },
                "answer": "Gorilla",
                # Below are optional
                "options": ["Dog", "Gorilla", "Hamster"]
            }
        ],
        # Will fetch `k` most relevant few-shot examples using embedding-based retriever
        k=2,
        # Lambda to turn the pd.DataFrame to a serialized string
        context_formatter=lambda df: df.to_markdown(
            index=False
        )
    )
}
smoothie = blend(
    query=blendsql,
    db=db,
    ingredients=ingredients,
    default_model=model,
)

LLMJoin

• Default examples
• All possible fields

from blendsql import blend, LLMJoin
from blendsql.ingredients.builtin import DEFAULT_JOIN_FEW_SHOT

ingredients = {
    LLMJoin.from_args(
        few_shot_examples=[
            *DEFAULT_JOIN_FEW_SHOT,
            {
                "join_criteria": "Join the state to its capital.",
                "left_values": ["California", "Massachusetts", "North Carolina"],
                "right_values": ["Sacramento", "Boston", "Chicago"],
                "mapping": {
                    "California": "Sacramento",
                    "Massachusetts": "Boston",
                    "North Carolina": "-"
                }
            }
        ],
        # Will fetch `k` most relevant few-shot examples using embedding-based retriever
        k=2
    )
}
smoothie = blend(
    query=blendsql,
    db=db,
    ingredients=ingredients,
    default_model=model,
)

Acknowledgements

Special thanks to those below for inspiring this project. Definitely recommend checking out the linked work below, and citing when applicable!

• The authors of Binding Language Models in Symbolic Languages
  • This paper was the primary inspiration for BlendSQL.
• The authors of EHRXQA: A Multi-Modal Question Answering Dataset for Electronic Health Records with Chest X-ray Images
  • As far as I can tell, the first publication to propose unifying model calls within SQL
  • Served as the inspiration for the vqa-ingredient.ipynb example
• The authors of Grammar Prompting for Domain-Specific Language Generation with Large Language Models
• The maintainers of the Guidance library for powering the constrained decoding capabilities of BlendSQL