2-Building Multi Agentic AI RAG With Vector Database

Agentic AI can be made to answer questions by pulling knowledge from a vector database that’s populated from PDFs—turning raw documents into a searchable “knowledge base” the assistant can query on demand. The core build here wires an assistant to a local PG Vector instance running in Docker, then loads PDF content (via a URL) into vector embeddings so the assistant can retrieve relevant passages and generate grounded responses.

The workflow starts with a practical fix: when using Gro-based setups, the code must explicitly provide a Gro API key rather than relying on defaults that may fall back to OpenAI. With that environment configuration in place, the project shifts to infrastructure—running PG Vector locally through Docker Desktop. Once the database is up, the system uses a “knowledge base” layer that accepts one or more PDF URLs, extracts text from those PDFs, converts the text into vector embeddings, and stores them in a named collection inside PG Vector.

From there, the assistant is created as a function (named consistently with the configuration) and connected to three capabilities: search the knowledge base, read chat history, and expose tools in responses. The assistant is configured with a run ID (initially none, then assigned after the first start), a user identifier, and the knowledge base object. Key toggles—such as enabling knowledge search and chat-history reading—allow the assistant to both retrieve relevant document chunks and maintain conversational context.

A concrete example uses a recipe PDF hosted on Amazon S3 (the URL points to a “recipes.pdf” document). After the program runs, it reports that documents have been added to the vector database. Then the assistant answers questions like ingredients and preparation steps—for example listing ingredient quantities (e.g., chicken, roasted peanuts) and returning directions for making specific dishes. The accuracy comes from retrieval: responses are generated using the most relevant chunks from the embedded PDF content stored in PG Vector.

Implementation details matter because the build depends on several libraries. The setup includes installing dependencies such as SQLAlchemy, PG Vector, psycopg binary, and PyPDF for PDF reading. The project also emphasizes that the same pattern can be repeated with other vector databases (the discussion names alternatives like Qdrant, Pinecone, LanceDB, ChromaDB, and SingleStore), but the walkthrough focuses on PG Vector and PG Vector 2.

The takeaway is less about a single chatbot and more about building an end-to-end agentic RAG pipeline: Docker-hosted vector storage → PDF ingestion via URL → embeddings + collection creation → assistant configured to search and respond. The assignment at the end pushes the same pipeline into a user-facing app using Streamlit, turning the backend workflow into an interactive front end for inputs and chat-style answers.