AIchivist

Search Washington State University's archival collections using natural language, powered by AI.

The Problem

University archives hold thousands of collections — letters, photographs, research papers, institutional records — but finding relevant material means wrestling with rigid keyword searches across catalog systems designed for librarians, not researchers. You have to already know the right terminology to find what you're looking for.

AIchivist lets you search the way you think. Ask a question in plain English, and the system combines AI query expansion with full-text search to surface the most relevant collections, ranked and explained.

Quick Start

You'll need .NET 10 SDK, Node.js, and an Anthropic API key.

# 1. Set your API key
dotnet user-secrets set "ANTHROPIC_API_KEY" "sk-ant-..." --project backend/ArchiveSearch.API

# 2. Start the backend (auto-creates SQLite database and applies migrations)
dotnet run --project backend/ArchiveSearch.API

# 3. In a new terminal — start the frontend
cd frontend && npm install && npm start

Open http://localhost:4200. Try searching: "What materials exist about student protests in the 1970s?"

How It Works

AIchivist uses a 3-pass search pipeline to go from a natural-language question to ranked, explained results:

User Query
    │
    ▼
┌─────────────────────────────────────┐
│  Pass 1: Query Expansion            │
│  Claude Haiku expands your query    │
│  into 6-8 synonyms & related terms  │
│  Detects date ranges automatically  │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│  Pass 2: Full-Text Search           │
│  SQLite FTS5 searches the catalog   │
│  using original + expanded terms    │
│  Weighted by title, abstract,       │
│  subjects, names, places            │
│  → ~30-50 candidate collections     │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│  Pass 3: AI Ranking                 │
│  Claude Haiku reads full metadata   │
│  for all candidates, selects and    │
│  ranks top 10 with explanations     │
│  → Scored 1-10 with relevance notes │
└─────────────────────────────────────┘

Each pass has a fallback — if query expansion fails, the original query is used. If ranking fails, results fall back to search-engine ordering. The pipeline degrades gracefully instead of breaking.

Results are cached for 1 hour, so repeated searches are instant.

Features

• Natural-language search — ask questions the way you'd ask a librarian, not a database
• AI-ranked results — top 10 results scored and explained by relevance, not just keyword frequency
• Follow-up chat — ask questions about your results in a chat sidebar, powered by Claude Sonnet
• Related collections — discover similar collections based on shared subjects, people, and places
• Detailed metadata — expand any result to see abstracts, scope notes, biographical history, series info, and more
• One-time setup — first-run wizard walks you through API key configuration with no terminal required

Tech Stack

Layer	Technology
Frontend	Angular 21 (zoneless, signal-based), SCSS, Vitest
Backend	ASP.NET Core (.NET 10), C#
Database	SQLite (FTS5), Entity Framework Core
AI	Anthropic Claude — Haiku 4.5 (search), Sonnet 4.5 (chat)

Commands

Backend

dotnet build backend/ArchiveSearch.slnx         # Build
dotnet run --project backend/ArchiveSearch.API   # Run API on http://localhost:5265
dotnet test backend/                             # Run tests

Frontend (from frontend/)

npm start       # Dev server on http://localhost:4200
npm run build   # Production build
npm test        # Run Vitest tests

Database

The SQLite database file (archive.db) is created automatically on first run -- no setup needed. To reset the database, simply delete the archive.db file and restart the backend.

API Endpoints

Endpoint	Method	Purpose
/api/search	POST	Run a search query
/api/chat	POST	Chat about search results
/api/admin/index	POST	Index EAD XML finding aids
/api/health	GET	Health check
/api/setup/status	GET	Check if API key is configured
/api/setup/save	POST	Save API key (first-run)

Project Structure

backend/
  ArchiveSearch.API/         # Controllers, services, DI config
    Services/
      SearchService.cs       # 3-pass search pipeline orchestration
      ClaudeService.cs       # Anthropic SDK wrapper (expand, rank, chat)
  ArchiveSearch.Core/        # Models, EAD XML parser, search cache
  ArchiveSearch.Data/        # DbContext, repository (FTS queries), migrations
  ArchiveSearch.Tests/       # Integration tests (xUnit + WebApplicationFactory)

frontend/src/app/
  components/
    home/                    # Main layout — search + results + chat
    search-bar/              # Search input with clickable example queries
    results-panel/           # Result list with expandable details
      result-card/           # Individual result — score, metadata, related collections
    chat-sidebar/            # Multi-turn follow-up chat
    setup/                   # First-run API key configuration
  services/                  # HTTP clients for search, chat
  models/                    # TypeScript interfaces

Desktop Installer

AIchivist can be packaged as a standalone Windows application with an embedded SQLite database -- no Docker, no terminal, no technical setup required.

.\installer\build-installer.ps1    # Builds frontend → runs tests → publishes → compiles installer

The installer bundles everything into a single AIchivist-Setup.exe that handles database initialization and configuration automatically. See installer/ for details.

Contributing

Contributions are welcome. Please open an issue to discuss what you'd like to change before submitting a pull request.