AIBraries

How AI is Being Used in Discovery Platforms

Basic Anatomy of a RAG-bot

LLM

Basic Anatomy of a RAG-bot

Fine-tuning (RLHF)

LLM

Basic Anatomy of a RAG-bot

Prompt

LLM

Fine-tuning (RLHF)

"You are a helpful librarian. Answer the following patron question about the UPenn Libraries. QUESTION: {{ question }}"

Basic Anatomy of a RAG-bot

Context

LLM

Fine-tuning (RLHF)

"You are a helpful librarian. Answer the following patron question about the UPenn Libraries using only information in the provided context. QUESTION: {{ question }} CONTEXT: {{context}}"

Prompt

"Van Pelt is open 8:30am - 6pm"

List all articles about the Soviet dissident movement written in the last five years but not authored by Ben Nathans

I don’t have real-time access to databases or the internet to provide the latest articles...

Basic Anatomy of a RAG-bot

Context

LLM

Fine-tuning (RLHF)

Prompt

knowledge base

Retrieval Augmented Generation (RAG)

• Texts are cut into chunks equal to the LLM's context size
• Text chunks are then transformed into numerical representations called text embeddings.
• Text embeddings that are similar to the question are retrieved and added as context to the prompt.

"Van Pelt is open 8:30am - 6pm"

"White Dog opens at 11 am"

"The US Open is coming to Philadelphia"

"The library's main entrances is closed." 

When is the library open?

[0.332,0.774,0.563 ...]

RAG demo

huggingface.co/spaces/prithivMLmods/RAG-PDF-CHATBOT

Use any PDF, or if you don't have one

• Benefits Enrollment Guide
• Collective Bargaining Agreement

List all articles about the Soviet dissident movement written in the last five years but not authored by Ben Nathans

Chain of Thought

"You are a helpful UPenn library research assistant. Think step-by-step and help the user solve their research problem." 

To solve the problem.
* First, I would go to franklin.library.upenn.edu
* Select the articles tab 
* Then I would enter "Soviet dissident movement"
* adjust the Publication Date filter...

Agents & Tools

Franklin Tool

Search Agent

"I need to search 'Soviet dissidents' where type is article and date is greater than 2020"

results = { "title":"Soviet Dissidents and their Friends","author","Fred Tyming"}

Serp Tool

"Google Scholar search..."

 Chain

 Graph

Discussion

Where have you seen similar chat-based products?

What information in needed from vendors for librarians to support these new systems?

What new capabilities do agentic systems bring to research and discovery?

How do you think chat-based interfaces will change research

practices?