Advanced AI: RAG, MCP, and Agents

Data Literacy

AI Workflows

• Process information from unstructured to structured data.
• Finding Aid for complex information.
• RAG (Retrieval-Augmented Generation)
• MCP (Multi-Chain Prompting)
• Agents (AI Agents)

RAG: Retrieval-Augmented Generation

• Process Data.
• Chunking.
• Embeddings.
• Vector Databases.
• LLMs (Large Language Models).

MCP: Model Context Protocol

• A standard protocol.
• Set of instructions for LLMs to follow.
• Gives LLM way to use external tools.
• LLMs can use multiple tools in a single prompt.
• Access third-party APIs.
• How to query databases.

Agents: AI Agents

• LLMs can act as agents.
• Agents can perform tasks on behalf of users.
• Agents can use tools and APIs to complete tasks.

Agents: AI Agents (Continued)

• Autonomously complete task at predetermined interval:
  • Check for new data.
  • Process data.
  • Generate report.
• Big Agents = Bad Agents

Breaking down a Workflow

• Break down a workflow into steps.
• Ask AI to do small jobs.
• Use deterministic scripts when possible.

Example 1: Indexing an Archive

• Step 1: Take Photos of Documents.
  • Human can include narrative to contextualize document.
• Step 2: Convert Photos to Text (OCR).
• Step 3: Chunk Text into Sections, if necessary.
• Step 4: Agent Extracts Metadata from Text.

Example 1: Indexing an Archive (Continued)

• Step 5: Agent Summarizes Text and Metadata.
• Step 6: Parse Metadata and Summaries into structured data.
• Step 7: Store structured data in a database.
• Step 8: Human reviews and corrects structured data.

Example 2: Vectorized Database

• Step 1: Collect sources.
• Step 2: OCR sources.
• Step 3: Chunk text into sections.
• Step 4: Create embeddings for each chunk.
• Step 5: Store embeddings in a vector database along with Metadata.
• Step 6: Use RAG to query vector database.