Skip to content

episode15 - I Built an AI Agent in C# (Azure OpenAI + Qdrant)

Choose a tag to compare

View all tags
@sandeeppatil-newvision sandeeppatil-newvision released this 27 Oct 07:18
· 8 commits to series/dotnet-ai since this release
ep15
e8e3b6c

EP15: AI Agent with Tool Calling (Phase 3, Episode 4)

Watch the episode: https://www.youtube.com/watch?v=p7XXdovsRAc&t
Builds on: ep14 (AI Document Analyzer). Download this release so your code matches the video.

WHAT'S NEW

Backend (DotNetAI.API)

  • Services/AgentTools.cs
    • 3 tools: list_documents, search_documents, calculate
    • Tool definitions (name, description, JSON schema) sent to the model
    • ExecuteAsync runs the tool the model asks for
    • Calculate accepts only digits and + - * / ( ) . and spaces, then evaluates with DataTable.Compute
  • Services/AgentService.cs
    • RunAsync: streaming agent loop, max 6 steps
    • System prompt: search before answering, cite file and page, use calculate for maths, say "I couldn't find that in your documents"
    • ToolCallAccumulator rebuilds streamed tool calls
    • Tool errors are sent back to the model so it can recover
    • AgentMessage and AgentEvent records (tool_call, tool_result, token)
  • Services/DocumentService.cs
    • New SearchAllAsync: search all PDFs, or one by file name
    • New DocHit record: file name, page, text, score
  • Controllers/AgentV1Controller.cs
    • POST api/AgentV1/stream (Server-Sent Events)
    • Streams tool calls, tool results and tokens, then [DONE]
  • Program.cs: AgentTools and AgentService registered as singletons

Frontend (dotnetai-chat)

  • src/api/agentStream.js: reads the SSE stream chunk by chunk
  • AgentPage.jsx and Agent.css: chat with a live tool trace and a Stop button
  • Root.jsx: new Agent tab (all tabs keep their state)

RUN IT

  1. Start Qdrant in Docker (ports 6333 and 6334)
  2. Run DotNetAI.API (https://localhost:7036)
  3. In dotnetai-chat: npm install, then npm run dev
  4. Open http://localhost:5173, upload a PDF in the Documents tab
  5. Ask in the Agent tab, for example:
    "I travel for 5 nights. What's the most I can spend on hotel plus 5 days of meals?"

REQUIREMENTS

  • .NET 10 SDK, Node.js
  • Azure OpenAI deployments: gpt-4.1 (chat) and text-embedding-ada-002 (embeddings)
  • Qdrant (Docker)
  • ep14 code (DocumentService, Documents tab)

NOTES

  • The agent can only use read-only tools. EP16 attacks this agent and hardens it.
  • Always check an AI answer against the source PDF. The agent cites the file and page.
  • The route is api/AgentV1/stream. Using api/Agent/stream returns 404.

NEXT: EP16 AI Red Teaming

Assets 2
Loading