Constrained Sampling from Large Language Models: Producing Structured Output

Large Language Models (LLMs) are typically used to generate free-form text. However, many applications require output with specific structures or formats. This presentation introduces the concept of constrained sampling from LLMs, a method to produce output that adheres to predefined formats or structures. The key points we will cover:

​Traditional LLM usage vs. constrained sampling

​Examples of interesting applications like classification/regression and producing JSON

Overview of how it works

1. 1 | © Copyright 2024 Zilliz1 1| © Copyright 10/22/23 Zilliz 1| © Copyright 2024 Zilliz
Stefan Webb
Developer Advocate, Zilliz
[email protected]
https://www.linkedin.com/in/stefan-webb
https://x.com/stefan_webb
Unstructured Data Meetup | Host

2. 2 | © Copyright 2024 Zilliz2 2| © Copyright 10/22/23 Zilliz 2| © Copyright 2024 Zilliz
Welcome Speakers
Constrained Sampling
from LLMs: Producing
Structured Output
Infinite World Generation
with Outlines
BAML Cheaper, faster, and
more accurate function
calling - without any
fine-tuning
TECH TALK 1 TECH TALK 2 TECH TALK 3
Stefan Webb
DevRel, Zilliz
Cameron Pfiffer
Developer Relations Engineer, .txt
Vaibhav Gupta
CEO, BoundaryML

3. 3 | © Copyright 2024 Zilliz3
Star Milvus
for a chance
to win a prize
tonight!

4. 4 | © Copyright Zilliz4
01
Introduction to Milvus
and Vector Databases

5. 5 | © Copyright Zilliz5

6. 6 | © Copyright Zilliz6
Why is Semantic Search Difficult?

7. 7 | © Copyright Zilliz7





or
Apple





or
Rising dough





or
Change car tire
Rising Dough
Proofing Bread
✔
❌
Semantic Search vs. Lexical Search

8. 8 | © Copyright Zilliz8
Why is Semantic Search Important?
10%
Other
newly generated data in 2025
will be unstructured data90%
Data Source: The Digitization of the World by IDC

9. 9 | © Copyright Zilliz9
Solution: Deep Learning
Similarity Search

10. 10 | © Copyright Zilliz10
Embeddings and Vector Spaces

11. 11 | © Copyright Zilliz11
New Challenge: Search in Vector Spaces
How to Index and
Search?

●High-dimensional
●> 1000 dims
How to Scale?

●10-100 million vectors?
●Billions?
●Trillions?
●Billions of users?
Multiple Data Types?

●Text
●Images
●Audio
●Graphs
●…

12. 12 | © Copyright Zilliz12
About Milvus
Milvus is an open-source vector database for
GenAI projects. pip install on your laptop, plug into
popular AI dev tools, and push to production with
a single line of code.
30K
GitHub Stars
66M
Docker Pulls
400
Contributors
2.7K
Forks
Easy Setup
Pip-install to start coding in a notebook within seconds
Integration
Plug into OpenAI, Langchain, LlmaIndex, and many more
Reusable Code
Write once, and deploy with one line of code into the production
environment
Feature-rich
Dense & sparse embeddings, filtering, reranking and beyond

13. 13 | © Copyright 2024 Zilliz13
Milvus Users

14. 14 | © Copyright 2024 Zilliz14 14| © Copyright 10/22/23 Zilliz 14| © Copyright 2024 Zilliz
github.com/milvus-io/milvus

zilliz.com/learn/generative-ai

15. 15 | © Copyright Zilliz15
02
Constrained Sampling from LLMs

16. 16 | © Copyright Zilliz16
Basic Multinomial Sampling from LLMs
Willard and Louf, 2023

17. 17 | © Copyright Zilliz17
chessboard
(environment)
context
move
(action)
image instruction
JSON
instruction
code
Agents
Information
Extraction
Coding
Assistant
Classification /
Regression
question domain
class or number

18. 18 | © Copyright Zilliz18
Structured Output by Prompting
“Extract the main details from the following
resume and format your answer as JSON.
Please make it well-formatted - Iʼd be ever so
grateful. Thanks, ta!ˮ
Key Idea: Instruct the model to output a given format.

19. 19 | © Copyright Zilliz19
Structured Output by Guided Sampling
Key Idea: Mask logits of tokens that are not valid next tokens
Willard and Louf, 2023

20. 20 | © Copyright Zilliz20
One Implementation: Finite State Machine
RegEx
r’^\d*(\.\d+)?$’
Key Idea:
●Build FSM of valid strings
●Index valid next tokens for each state in machine.
●Run FSM while generating output, using valid next tokens
to produce mask on logits.
Mask(state)

21. 21 | © Copyright Zilliz21
Simple Example
Sampled: “ˮ
State: 0
Sampled: “.2ˮ
State: 3
Sampled: “1ˮ
State: 1
RegEx / Grammar:
“All numbers with only digits
and decimal point.ˮ

Tokens (vocab.):
“Aˮ, “.ˮ, “42ˮ, “.2ˮ, “1ˮ

22. 22 | © Copyright Zilliz22
Outlines Library (from .txt)

23. 23 | © Copyright Zilliz23
Outlines Library (from .txt)

24. 24 | © Copyright Zilliz24
Try it Out Yourself!
https://github.com/boundaryml/baml

https://github.com/dottxt-ai/outlines

25. 25 | © Copyright Zilliz25
T H A N K Y O U

26. 26 | © Copyright 2024 Zilliz26
https://milvus.io/discord
https://github.com/milvus-io/milvus
https://x.com/milvusio
https://www.linkedin.com/company/the-milvus-project
LET’S STAY CONNECTED!
Stefan Webb
Developer Advocate, Zilliz

27. 27 | © Copyright 2024 Zilliz27
Become a
Speaker!
Interesting in speaking
at and/or sponsoring a
Zilliz Unstructured Data Meetup?
Fill out this form


??????????????????

28. 28 | © Copyright 2024 Zilliz28
Join us at our next meetup!
lu.ma/unstructured-data-meetup
Nov 13, South Bay, Location TBD
“Evaluating Retrieval-Augmented Generationˮ

Nov 19, San Francisco, GitHub Office
“Structured Output from Unstructured Input: Part 2ˮ