1 of 27

The Future of Knowledge Assistants

Jerry Liu

July 9th, 2024

2 of 27

Building a Knowledge Assistant

2

Agent: …

Human: …

Agent: …

Knowledge Base

Answer:

Sources: …

Human: <Question>

Human: …

Goal: Build an interface that can take in any task as input and give back an output.

Input forms: simple questions, complex questions, research tasks

Output forms: short answer, structured output, research report

3 of 27

The LLM Application Ecosystem

Enterprise developer teams are the best equipped to build a new category of AI-powered software.

But ⚠️

Developers need tooling and infrastructure to efficiently build production-quality LLM applications.

Data

Foundation Models

Knowledge-Intensive LLM Applications

Sales

Dev

Legal

Finance

🛠️ Tooling + Infra

🧑‍💻 Developers

4 of 27

LlamaIndex: Build Production LLM Apps over Enterprise Data

LlamaIndex helps any developer build context-augmented LLM apps from prototype to production.

Open-Source: Leading developer toolkit for building production LLM apps over data.

1M+ monthly downloads

170K+ LinkedIn followers

30k+ stars

LlamaCloud: A centralized knowledge interface for your production LLM application.

5 of 27

RAG was just the beginning

🧠 🛠️

6 of 27

Knowledge Assistant with Basic RAG

⚠️ Naive data processing, primitive retrieval interface

⚠️ Poor query understanding/planning

⚠️ No function calling or tool use

⚠️ Stateless, no memory

6

Data Processing and Indexing

Index

Data

Basic Text Splitting

Top-k = 5

Simple QA Prompt

Response

Basic Retrieval and Prompting

7 of 27

Problem: Productionizing LLM Applications at Scale

Pain Points

Accuracy Issues:

Bad retrievals, hallucinations

Hard to Improve:

Too many parameters, deep-tech expertise required

Hard to Scale:

every new data source requires eng hours, custom parsing, tuning

Prototype

Productionize

Small set of simple documents

Low quality bar

Large set of complex, heterogeneous documents

High quality bar

8 of 27

Data Silos Exacerbate this Problem

Structured data sits in Snowflake, BigQuery, and Redshift clusters

Unstructured data sits in your S3 buckets, filesystems

APIs

Raw Files

SQL DBs

Vector Stores

Production LLM Apps

Q&A

Chat

Copilot

Autonomous Agents

Structured Extraction

?

?

?

?

9 of 27

Can we do more?

There’s many questions/tasks that naive RAG can’t give an answer to.

🚫 Hallucinations

🚫 Limited time savings

🚫 Limited decision-making enhancement

9

💡 How do we aim to build a production-ready knowledge assistant?

10 of 27

A Better Knowledge Assistant

  1. Advanced data and retrieval modules
  2. Advanced single-agent query flows
  3. General multi-agent task solvers

10

Advanced Data and Retrieval Tool

Index

Data

Data Processing

Agent

Response

11 of 27

A Better Knowledge Assistant

  • Advanced data and retrieval modules
  • Advanced single-agent query flows
  • General multi-agent task solvers

11

Advanced RAG and Retrieval Tool

Response

Other tools

Other tools

Agent

Memory

Tool Use

Query Planning

Reflection

12 of 27

A Better Knowledge Assistant

  • Advanced data and retrieval modules
  • Advanced single-agent query flows
  • General multi-agent task solvers

12

Agent 1

E.g. Orchestrator

Agent 2

E.g. RAG

Agent 3

E.g. SQL

Task

Agent 4

E.g. Reflection

Response

13 of 27

Advanced Data and Retrieval

🧠 🛠️

14 of 27

Any LLM App is only as Good as your Data

Garbage in = garbage out

Good data quality is a necessary component of any production LLM app.

14

Raw Data

Data Processing

Clean Data

Production LLM Apps

Q&A

Chat

Copilot

Autonomous Agents

Structured Extraction

15 of 27

Any LLM App is only as Good as your Data

15

ETL for LLMs

  • Parsing
  • Chunking
  • Indexing

Index

Data

Data Processing

Agent

Response

16 of 27

LlamaCloud: Parsing (LlamaParse)

  • Advanced document parser specifically for reducing LLM hallucinations
    • ~500k+ monthly downloads
    • 14k+ unique users
    • 13M+ pages processed

“As an AI Applied Data Scientist who was granted one of the first ML patents in the U.S., and who is building cutting-edge AI capabilities at one of the world’s largest Private Equity Funds, I can confidently say that LlamaParse from LlamaIndex is currently the best technology I have seen for parsing complex document structures for Enterprise RAG pipelines. Its ability to preserve nested tables, extract challenging spatial layouts, and images is key to maintaining data integrity in advanced RAG and agentic model building.”

Dean Barr, Applied AI Lead at Carlyle

17 of 27

LlamaCloud: Advanced ETL

Centralize and enhance your data for your production LLM application.

  • Enterprise-Ready Data Connectors
  • Complex Document Parsing (LlamaParse)
  • Continuous Synchronization
  • Enterprise-Ready Data Sinks
  • Advanced Retrieval
  • Retrieval observability and evaluations

“LlamaCloud has really sped up our development timelines. Getting to technical prototypes quickly allows us to show tangible value instantly, improving our sales outcomes. When needed, switching from the LlamaCloud UI to code has been really seamless. The configurable parsing and retrieval features have significantly improved our response accuracy. We've also seen great results with LlamaParse and found it outperforming GPT-4 vision on some OCR tasks!”

Teemu Lahdenpera, CTO at Scaleport.ai

18 of 27

Use Cases

Customers (pending name approval):

  • Fortune 500 Investment Firm
  • Big 4 Consulting Firm
  • Global Investment Bank
  • Japanese Technology Conglomerate
  • Leading IoT Supplier for Supply Chain Tracking
  • Global Health + AI Company

Select Verticals

financial services, consulting, healthcare, tech

Select Use Cases

Financial Analyst Assistant

Centralized Internal Search (Organization + User-level)

Analytics Dashboard for Sensor Data

Internal LLM Application Development Platform

19 of 27

Advanced Single-Agent Flows

🧠 🛠️

20 of 27

From Simple to Advanced Agents

20

Full Agents

Agent Ingredients

Simple

Lower Cost

Lower Latency

Advanced

Higher Cost

Higher Latency

Routing

One-Shot Query Planning

Tool Use

ReAct

Dynamic Planning + Execution

Conversation Memory

21 of 27

Agentic RAG

Every data interface is a tool

Use agent reasoning loops (sequential, DAG, tree) to tackle complex tasks

End Result: Build personalized QA systems capable of handling complex questions!

21

21

Advanced RAG and Retrieval Tool

Response

Other tools

Other tools

Agent

Memory

Tool Use

Query Planning

Reflection

22 of 27

Remaining Gaps

A single agent cannot solve an infinite set of tasks - specialist agents do better!

Agents are increasingly interfacing with services that may be other agents

22

23 of 27

Multi-Agent Task Solvers

🧠 🛠️

24 of 27

Multi-Agents: Motivation and Use Case

Why Multi-Agents?

  • Specialization
  • Parallelization
  • Cost/Latency

Challenges:

  • Adding guardrails
  • Converting notebook code to production

24

25 of 27

Llama Agents (Preview)

Agents as microservices.

  • Encapsulation and modularity
  • Communication via standardized API interfaces
  • Ease of deployment
  • Scalability and resource management

25

26 of 27

Llama Agents: Architecture

26

27 of 27

LlamaCloud: Opening up a Waitlist

Advanced unstructured data processing for LLMs

27

Advanced ETL

Advanced Parsing

Scan here