This repository demonstrates using Ollama (local LLM) with Microsoft Agent Framework to summarize PDF documents and generate embeddings from documents for future use. It is Based on PDF_Llama. Uses Microsoft Agent Framework.

Usage scenarios

• Analyse organisations PDF documents and create embeddings for search.
• Document content summarisations.
• Rapid insights of large PDF collections. Get informartion of individiual documents and sort PDF collections.

Architecture

Uses plugin pattern where the Kernel is injected into the plugin methods when required. Allows the plugin to use the kernel's AI services as configured. PDFs read using PDFPig. Besides textual content, this helpful piggy gives us the ability to AI analyse documents images, links and metadata.

Working on atm...

• Using gemma3:4b model to add image analysis
• Serach for 'gemma3'

Choosing an embedding model

• Model accuracy and dimensionality.
• What is the size of the vectors generated by the model,
• Is it configurable, as this will affect your vector storage cost.
• What type of elements does the generated vectors contain, e.g. float32, float16 - cost again.
• How fast does it generate vectors?

Future enhancements

User definable analysis steps

Offer logical steps for users who create sets for analysis. Steps can be defined using function invocation filters. Users are able to opt in/out of certain analysis steps based on documents type or content. For exsample:

• AI is instructed to exclude personal data from medical documents.
• Images are analysed for certain document types, such as marketing materials that contains certain product/object.

Analysis sets

Sets could be for finacial analysis, for organisations task history. Set of product information, set to contain historical price fluctuations.

Add Aspire MCP

Use Agents https://devblogs.microsoft.com/aspire/scaling-ai-agents-with-aspire-isolation.

Prerequisites

• .NET 10 SDK
• Ollama installed and running locally (default endpoint http://localhost:11434).
  • Download model, llama3.2 works fine.
  • To download: ollama pull llama3.2
• Ollama access, work in progress... Will decide between NuGet packages:
  • OllamaSharp or Microsoft.AI.Extensions.Ollama
• Process by
  • Microsoft.SemanticKernel.

Install packages using the CLI:

dotnet add package OllamaSharp
dotnet add package Microsoft.SemanticKernel

How to run

1. Start Ollama: follow Ollama docs to run the service locally.
2. Build the project:

dotnet build

3. Run the sample that uses the summariser:

dotnet run --project PDF_Llama.csproj

Footnotes

• If you intentionally want to use an alternate Microsoft extension package, ensure you add the correct NuGet package that provides Microsoft.Extensions.AI.Ollama or update code to match available libraries.

Common fix for CS0234

You may see an error:

CS0234: The type or namespace name 'Ollama' does not exist in the namespace 'Microsoft.Extensions.AI' (are you missing an assembly reference?)

This usually means the code references a namespace that isn't provided by any installed package. In this project use OllamaSharp types and remove using Microsoft.Extensions.AI.Ollama; unless you add a package that actually provides that namespace.

Steps to resolve:

1. Remove the incorrect using directive from files (for example DotNetAI.cs and others):

   • using Microsoft.Extensions.AI.Ollama; -> remove it

2. Use the OllamaSharp namespace: ensure files include using OllamaSharp; and using OllamaSharp.Models; where needed.

3. Install the OllamaSharp NuGet package: dotnet add package OllamaSharp and restore packages.

4. Ensure Ollama is running and the expected model is downloaded, e.g. llama3.2.

    // https://stackoverflow.com/questions/1653046/what-are-the-true-benefits-of-expandoobject
    // https://www.daveabrock.com/2021/01/19/config-top-level-programs/