Minutes Studio

A prototype Retrieval-Augmented Generation (RAG) application that ingests meeting-minute PDFs and generates grounded, cited analyst work products — plus a free-form document chat — using Azure OpenAI and Azure AI Search.

It is purpose-built for meeting minutes from congressional committee meetings — the domain shapes the whole app: work products are the kinds of deliverables a committee/legislative-affairs team produces (stakeholder briefs, executive talking points, plain-language summaries), and bills and nominations mentioned in the minutes (e.g. S.3018, H.R.260, PN893) are automatically linked to congress.gov, resolved to the correct Congress from the meeting date.

Everything the model produces is grounded in the source documents and cited back to the specific meeting (and page range), so answers stay traceable and free of fabrication.

• Work Products — one-click generation from selected documents:
  • Stakeholder Brief
  • Executive Talking Points
  • Plain-Language Summary
  • Meeting Summary
• Document Chat — free-form Q&A across all documents (or scoped to one), with hybrid retrieval.
• Single vs. all meetings — single-meeting generation uses the document's full text; "all meetings" uses a map-reduce pipeline to avoid cross-meeting fact bleed.
• Grounded citations — click a citation to open a side panel with the source excerpt.
• Congress.gov linking — detected bills/nominations (e.g. S.3018, PN893) are linked to congress.gov, resolved to the correct Congress from the meeting date, and only linked when grounded in the source.
• Tone / length / reference controls — adjust output style and choose whether references are included, hidden, or stripped for a clean, email-safe copy.
• Token accounting — input/output/total token usage surfaced per generation.
• Document management — list, preview (inline PDF), upload, and (re)index PDFs in Blob Storage.
• Health preflight — checks embeddings, chat, and search connectivity before you ingest.

• .NET 8, ASP.NET Core Blazor Web App (Interactive Server) — MinutesStudio.Web
• RAG core class library — MinutesStudio.Core
• Azure OpenAI (Azure AI Foundry) — gpt-5.4-mini (chat) + text-embedding-3-large (3072-dim embeddings)
• Azure AI Search — hybrid (BM25 + HNSW vector) index
• Azure Blob Storage — source PDFs
• UglyToad.PdfPig — PDF text extraction · Markdig — Markdown rendering

.
├── MinutesStudio.sln
├── docs/
│   └── data-architecture.md        # Detailed data flow + Mermaid diagrams
├── samples/                        # Sample meeting-minute PDFs
└── src/
    ├── MinutesStudio.Core/         # RAG logic (no UI dependencies)
    │   ├── Configuration/          # Options: AzureOpenAI, AzureSearch, AzureBlob, Rag
    │   ├── Models/
    │   ├── Prompts/                # PromptLibrary + grounding rules
    │   └── Services/               # Ingestion, embeddings, search, generation, RAG, etc.
    └── MinutesStudio.Web/          # Blazor UI + minimal JSON API
        ├── Components/             # Pages, layout, ReferencePanel
        └── wwwroot/

flowchart LR
    subgraph Source
        U[User / analyst]
        LF[Local samples folder]
    end
    subgraph Azure
        BLOB[(Blob Storage<br/>PDFs)]
        AOAI[Azure OpenAI<br/>embeddings + chat]
        SEARCH[(Azure AI Search<br/>hybrid index)]
    end
    subgraph App[Blazor Web App + MinutesStudio.Core]
        ING[IngestionService]
        RAG[RagService]
    end
    LF -->|Upload samples| BLOB
    U -->|Upload PDF| BLOB
    BLOB -->|list + stream| ING
    ING -->|embed chunks| AOAI
    ING -->|upload vectors| SEARCH
    U -->|Generate / Ask| RAG
    RAG -->|retrieve / full text| SEARCH
    RAG -->|embed query + complete| AOAI
    RAG -->|grounded, cited output| U

See docs/data-architecture.md for the full ingestion pipeline, map-reduce generation, index model, and resilience details. The doc is also viewable in-app on the Architecture page.

• .NET 8 SDK
• An Azure OpenAI (Azure AI Foundry) resource with two deployments:
  • a chat model (default deployment name gpt-5.4-mini)
  • text-embedding-3-large (embeddings)
• An Azure AI Search service (Basic tier or higher; needs an admin key or RBAC).
• An Azure Blob Storage account with a container for the source PDFs.

Non-secret defaults (deployment / index / container names, chunking) live in src/MinutesStudio.Web/appsettings.Development.json. Secrets are never committed — set them via .NET user-secrets.

cd src/MinutesStudio.Web

# Azure OpenAI (Foundry)
dotnet user-secrets set "AzureOpenAI:Endpoint" "https://<your-foundry>.services.ai.azure.com/"
dotnet user-secrets set "AzureOpenAI:ApiKey"   "<your-aoai-key>"

# Azure AI Search
dotnet user-secrets set "AzureSearch:Endpoint" "https://<your-search>.search.windows.net"
dotnet user-secrets set "AzureSearch:ApiKey"   "<your-search-admin-key>"

# Azure Blob Storage
dotnet user-secrets set "AzureBlob:ConnectionString" "<your-storage-connection-string>"

If an API key / connection string is left empty, the app falls back to DefaultAzureCredential (managed identity / az login), which is the preferred path for an Azure deployment.

cd src/MinutesStudio.Web
dotnet run

Then open http://localhost:5190 (the default http launch profile, which sets the Development environment so your user-secrets load).

1. Go to the Documents page.
2. Upload samples (seeds Blob Storage from the local samples/ folder) or upload your own PDFs.
3. Click Ingest to extract → chunk → embed → index the PDFs.
4. Head to Work Products or Chat and start generating.

The Azure AI Search index prefix is minutesstudio-minutes. Resetting the index creates a new timestamped index and cleans up the old ones automatically.

A lightweight JSON API is exposed for automation and scripting:

type accepts the work-product names (e.g. StakeholderBrief, ExecutiveTalkingPoints, PlainLanguageSummary, MeetingSummary). references accepts Included, Hidden, or Clean.

The four work-product prompts and the shared grounding rules (injected into every prompt and the chat assistant) are defined in MinutesStudio.Core/Prompts/PromptLibrary.cs and are viewable in-app on the Prompt Library page. Each template is annotated with the key design decisions behind it.

• Transient-fault retry with exponential backoff wraps embedding and chat calls (treats 408/429/5xx, network errors, and intermittent 404 DeploymentNotFound as transient).
• The Azure OpenAI client API version is pinned to 2024-10-21.
• SDK exceptions are mapped to actionable messages (auth / missing / throttled / transient).

• Entra ID authentication on the app + managed identity to Azure OpenAI, Search, and Blob (removing API keys entirely).
• Azure deployment (App Service / Container Apps).

Prototype built for evaluation. Not production-hardened — secrets are local user-secrets and the auth model is key-based pending the Entra/managed-identity work above.

Released under the MIT License.