A self-hosted platform for conducting AI-powered conversational interviews at scale.
Built for survey researchers, qualitative methodologists, and anyone who needs structured or open-ended data collection through voice or text.
Website · Docs · About · License
OASIS is an open-source research platform that lets you run conversational AI interviews from your own infrastructure. You define the study, configure an agent with a system prompt and model of your choice, and share a link (or phone number) with participants. Transcripts are diarized and stored in your database. You stay in control of the data, the models, and the entire pipeline.
It was built out of a simple frustration: existing commercial tools for conversational AI are powerful, but they are not designed with research methodology in mind. Things like follow-up probing, semi-structured interview guides, participant identifiers, and study-level organization tend to be afterthoughts. OASIS puts those features front and center.
OASIS launched publicly in March 2026. While we have tested the platform extensively across different virtual machines, operating systems, and deployment setups, it is still a young project. You may encounter rough edges, especially in less common configurations. If something breaks, please open an issue and we will do our best to help.
1. Setup a Study
demo1-fast.mp4
2. Create and Configure an Agent
demo2-fast.mp4
3. Collect and Manage Data
demo3-fast.mp4
Two interview modalities
- Voice interviews (real-time speech with STT, LLM, and TTS)
- Text chat interviews (clean chat UI with customizable avatars)
Flexible pipeline architecture
- Modular pipeline: chain any STT, LLM, and TTS provider independently
- Voice-to-voice pipeline: stream audio directly to multimodal models (OpenAI Realtime, Gemini Live)
Semi-structured interview mode
- Define a question guide with preset questions, follow-up probes, and transition logic
- The agent follows a structured backbone while still allowing natural conversation
- Upload interview guides via CSV or JSON, or configure them in the dashboard
Multi-provider model support
- OpenAI (GPT-4o, GPT-4.1, GPT-5, Realtime models)
- Google Gemini (including native audio preview)
- Scaleway (European-hosted, OpenAI-compatible)
- Azure OpenAI and GCP Vertex AI (self-hosted endpoints)
- Any LiteLLM-compatible provider, plus a custom model identifier option
STT and TTS providers
- Deepgram, Scaleway Whisper (STT)
- ElevenLabs, Cartesia, Scaleway (TTS)
Research-first design
- Study-level organization with multiple agents per study
- Participant identifiers: random, predefined lists, or self-reported
- Diarized transcripts with timestamps
- Session analytics and data export
- Knowledge base (RAG) with pgvector for document-grounded conversations
Telephony support (beta)
- Twilio Media Streams integration for phone-based interviews
- Currently supports incoming calls
Self-hosted and private
- Everything runs in Docker on your own machine or server
- No data leaves your infrastructure unless you explicitly configure external API calls
- Optional basic authentication for the admin dashboard
- API keys configurable via
.envor the dashboard settings page
A note on GPUs and HPCs
- OASIS does not currently include built-in support for GPU scheduling or HPC cluster integration. This is intentional. HPC environments vary significantly between institutions, and teams that manage GPU clusters typically have the technical expertise to integrate OASIS into their own infrastructure. If you are running OASIS on an HPC and want to connect it to local GPU-hosted models, the LiteLLM layer makes this straightforward via custom endpoints.
OASIS runs as five Docker containers on a single internal network:
| Container | Stack | Role |
|---|---|---|
| Caddy | Caddy 2 | Reverse proxy, automatic HTTPS |
| Frontend | React, Tailwind CSS, Nginx | Admin dashboard and participant interview widget |
| Backend | FastAPI, Pipecat, LiteLLM | WebSocket audio/text transport, AI pipeline orchestration, REST API |
| PostgreSQL | pgvector/pg16 | Agent configs, transcripts, participant data, vector embeddings |
| Redis | Redis 7 | Session tracking, real-time transcript pub/sub, API key overrides |
┌──────────────────────────────────────────────────────────┐
│ Caddy (ports 80/443) │
│ ┌──────────┬──────────┐ │
│ │ Frontend │ Backend │ │
│ │ (React) │ (FastAPI)│ │
│ └────┬─────┴────┬─────┘ │
│ │ │ │
│ ┌────┴────┐ ┌───┴───┐ │
│ │PostgreSQL│ │ Redis │ │
│ └─────────┘ └───────┘ │
└──────────────────────────────────────────────────────────┘
- Docker and Docker Compose
- At least one AI provider API key (OpenAI, Google, Scaleway, etc.)
git clone https://github.com/oasis-surveys/oasis-platform.git
cd oasis-platformCopy the example .env file and fill in your API keys:
cp .env.example .envAt a minimum, you will need:
# Required
OPENAI_API_KEY=sk-...
SECRET_KEY=some-random-secret-string
# Recommended for voice interviews
DEEPGRAM_API_KEY=...
ELEVENLABS_API_KEY=...
# Optional providers
GOOGLE_API_KEY=...
SCALEWAY_SECRET_KEY=...
SCALEWAY_PROJECT_ID=...
# Optional: enable login page
AUTH_ENABLED=false
AUTH_USERNAME=admin
AUTH_PASSWORD=your-password
# Optional: Twilio for phone interviews
TWILIO_ACCOUNT_SID=...
TWILIO_AUTH_TOKEN=...
TWILIO_PHONE_NUMBER=...docker compose up -dThat is it. Open your browser and go to http://localhost (or your server's address). The admin dashboard will be ready.
- Click "New Study" in the dashboard
- Add an agent, write a system prompt, select your models
- Set the agent to Active
- Copy the interview link and share it with participants
oasis/
├── backend/
│ ├── app/
│ │ ├── api/ # REST + WebSocket endpoints
│ │ ├── models/ # SQLAlchemy ORM models
│ │ ├── schemas/ # Pydantic request/response schemas
│ │ ├── pipeline/ # Pipecat pipeline runner and processors
│ │ ├── knowledge/ # RAG: chunking, embedding, retrieval
│ │ ├── config.py # Environment-based settings
│ │ └── main.py # FastAPI app entry point
│ ├── alembic/ # Database migrations
│ ├── tests/ # Pytest test suite
│ └── Dockerfile
├── frontend/
│ ├── src/
│ │ ├── pages/ # Dashboard and interview widget pages
│ │ ├── components/ # Shared UI components
│ │ ├── contexts/ # React contexts (auth)
│ │ └── lib/ # API client, constants, utilities
│ └── Dockerfile
├── docker/
│ └── Caddyfile # Reverse proxy config
├── docker-compose.yml
└── .env.example
OASIS has a comprehensive test suite that runs without any real API keys. All external calls are mocked.
# Backend tests (from the backend/ directory)
pip install -r requirements.txt
pytest tests/ -v --cov=app
# Frontend tests (from the frontend/ directory)
npm install
npx vitest runTests also run automatically on every push via GitHub Actions. A weekly scheduled run makes sure nothing has drifted.
All settings are managed through environment variables. See backend/app/config.py for the full list, or check the documentation for a detailed walkthrough.
Key options:
| Variable | Description | Default |
|---|---|---|
OPENAI_API_KEY |
OpenAI API key | (required for OpenAI models) |
DEEPGRAM_API_KEY |
Deepgram STT key | (required for modular voice pipeline) |
ELEVENLABS_API_KEY |
ElevenLabs TTS key | (required for modular voice pipeline) |
GOOGLE_API_KEY |
Google AI Studio key | (optional, for Gemini models) |
SCALEWAY_SECRET_KEY |
Scaleway API key | (optional, European-hosted models) |
AUTH_ENABLED |
Enable dashboard login | false |
AUTH_USERNAME |
Login username | admin |
AUTH_PASSWORD |
Login password | (must be set if auth enabled) |
OASIS exists to fill a gap. Commercial conversational AI platforms are built for customer support, sales, and marketing. They are good at what they do, but research has different requirements:
- Methodological control. Researchers need semi-structured guides, probing logic, and participant tracking built into the tool, not bolted on.
- Transparency. When you publish a study, reviewers and readers should be able to understand exactly what system was used, what models were called, and where data was stored.
- Affordability. Academic budgets are not enterprise budgets. Self-hosting with pay-as-you-go API keys is often the only realistic option.
- Data sovereignty. Especially in Europe, running your own infrastructure is not a nice-to-have. It is a compliance requirement.
The long-term goal is for OASIS to become a standard open-source tool that survey and qualitative researchers can rely on, the same way they rely on tools like Qualtrics for traditional surveys but with conversational AI as the medium.
Contributions are welcome. Please read CONTRIBUTING.md before opening a pull request.
- Bug reports: Open a bug report
- Feature requests: Suggest a feature
- General questions: max.lang@stx.ox.ac.uk
OASIS is released under the Open Non-Commercial Research License (ONCRL) v1.0.
You are free to use, modify, and share OASIS for non-commercial research and educational purposes. Use as a core component of a funded research project or grant application requires prior written approval from the copyright holder — this is not meant to be a barrier, it just helps us understand how the project is being used and ensure proper attribution.
For universities and research institutions that prefer not to self-host, the OASIS Project may offer the platform as a hosted or managed service. Contributed code may be included in such offerings (see section 3 and 5 of the license). If you have questions about this, please get in touch.
See the full LICENSE file for details.
If you use OASIS in your research, please cite it:
@software{lang2026oasis,
author = {Lang, Max M.},
title = {{OASIS}: Open Agentic Survey Interview System},
year = {2026},
url = {https://github.com/oasis-surveys/oasis-platform},
note = {Self-hosted platform for AI-powered conversational interviews},
doi = {10.5281/zenodo.19041570}
}