AI-Powered New Hire Knowledge Accelerator built on Azure OpenAI + pgvector.

OnboardIQ ingests your company's M365 content such as SharePoint documents, Teams meeting transcripts, and OneNote notebooks, then answers new-hire questions with grounded, source-cited responses.

onboardiq/
|-- app/                  # FastAPI backend
|   |-- api/routes.py     # REST endpoints
|   |-- config/           # pydantic-settings
|   |-- core/             # DB engine, ORM models
|   |-- ingestion/        # chunker, embedder, Graph API client
|   |-- migrations/       # Alembic + pgvector indexes/constraints
|   `-- query/            # retriever, answer generator
`-- frontend/             # React + Vite + Tailwind UI
    `-- src/
        |-- components/
        |   |-- ChatView.tsx
        |   |-- IngestView.tsx
        |   `-- Sidebar.tsx
        |-- api.ts
        `-- App.tsx

1. MS Graph content ingestion (graph_client.py) OAuth2 token retrieval is implemented, but SharePoint file enumeration, content download, and Teams transcript retrieval still fall back to synthetic demo data.

2. Background ingestion queue (routes.py /ingest) Ingestion now runs in-process for local development. For production-scale workloads, move it to ARQ, Celery, or Azure Queue Storage.

3. Prompt tuning (generator.py _SYSTEM_PROMPT) The generator now returns structured JSON confidence, but the prompt can still be improved with few-shot examples for stronger grounding quality.

• Python 3.11+
• uv package manager
• Node.js 18+ and pnpm
• PostgreSQL 15+ with the pgvector extension enabled
• Azure OpenAI resource with text-embedding-3-large and gpt-4o deployments

The easiest way to run PostgreSQL locally is via Docker:

docker run -d \
  --name onboardiq-pg \
  -e POSTGRES_USER=postgres \
  -e POSTGRES_PASSWORD=postgres \
  -e POSTGRES_DB=onboardiq \
  -p 5432:5432 \
  pgvector/pgvector:pg16

Run this from the project root:

uv sync

cp .env.example .env
# Edit .env with your credentials

Key variables:

uv run alembic -c app/alembic.ini upgrade head

This creates the tables, vector support, and the unique constraint used for chunk upserts.

uv run uvicorn app.main:app --reload

API: http://localhost:8000

Docs: http://localhost:8000/docs

Health: http://localhost:8000/api/health

cd frontend
pnpm install
pnpm dev

Frontend: http://localhost:3000

The Vite dev server proxies /api/* requests to http://localhost:8000.

1. Start PostgreSQL and make sure .env contains real values, not placeholders. At minimum, DATABASE_URL and the credentials for your selected LLM_PROVIDER must be valid.

2. Install dependencies:

uv sync

3. Apply migrations:

uv run alembic -c app/alembic.ini upgrade head

4. Start the API:

uv run uvicorn app.main:app --reload

5. Verify the server:

• http://localhost:8000/
• http://localhost:8000/api/health
• http://localhost:8000/docs

6. Test ingestion in Swagger with POST /api/ingest:

{
  "source_type": "sharepoint",
  "site_id": "demo-site"
}

7. Test querying in Swagger with POST /api/query:

{
  "question": "When is the benefits enrollment deadline for new hires?"
}

• If uv run alembic upgrade head fails with No 'script_location' key found in configuration, use uv run alembic -c app/alembic.ini upgrade head. This repo stores alembic.ini inside app/.

• If http://localhost:8000/health returns 404, use http://localhost:8000/api/health. The API router is mounted under /api.

• If POST /api/ingest returns 500, check the terminal running uvicorn. The most common causes are:

  • DATABASE_URL still points to placeholder values
  • Azure/OpenAI credentials are still placeholders
  • database migrations were not applied

MS Graph API / demo content
        |
        v
graph_client.py         (token flow + source loading)
        |
        v
chunker.py              (split into token-bounded chunks with overlap)
        |
        v
embedder.py             (embeddings + pgvector upsert)
        |
        v
PostgreSQL + pgvector   (store chunks + vector search index)

User question (POST /api/query)
        |
        v
embedder.embed_text()   (embed question)
        |
        v
retriever.search()      (cosine ANN search)
        |
        v
generator.generate()    (grounded answer + structured confidence)
        |
        v
QueryResponse           (answer, sources, confidence)

1. Implement SharePoint file enumeration and content extraction via Microsoft Graph.
2. Implement Teams transcript retrieval via Microsoft Graph.
3. Move /ingest to a background task queue such as ARQ.
4. Add authentication middleware (Azure AD JWT validation) for production use.
5. Write integration tests against a local pgvector Docker container.
6. Set up Azure Container Apps or AKS deployment with Helm charts.