---
title: Testing
description: Running and writing tests for SurfSense
---

SurfSense uses [pytest](https://docs.pytest.org/) with two test layers: **unit** tests (no database) and **integration** tests (require PostgreSQL + pgvector). Tests are self-bootstrapping — they configure the test database, register a user, and clean up automatically.

## Prerequisites

- **PostgreSQL + pgvector** running locally (database `surfsense_test` will be used)
- **`REGISTRATION_ENABLED=TRUE`** in your `.env` (this is the default)
- A working LLM model with a valid API key in `global_llm_config.yaml` (for integration tests)

No Redis or Celery is required — integration tests use an inline task dispatcher.

## Running Tests

**Run all tests:**

```bash
uv run pytest
```

**Run by marker:**

```bash
uv run pytest -m unit          # fast, no DB needed
uv run pytest -m integration   # requires PostgreSQL + pgvector
```

**Available markers:**

| Marker | Description |
|---|---|
| `unit` | Pure logic tests, no DB or external services |
| `integration` | Tests that require a real PostgreSQL database |

**Useful flags:**

| Flag | Description |
|---|---|
| `-s` | Show live output (useful for debugging polling loops) |
| `--tb=long` | Full tracebacks instead of short summaries |
| `-k "test_name"` | Run a single test by name |
| `-o addopts=""` | Override default flags from `pyproject.toml` |

## Configuration

Default pytest options are in `surfsense_backend/pyproject.toml`:

```toml
[tool.pytest.ini_options]
addopts = "-v --tb=short -x --strict-markers -ra --durations=5"
```

- `-v` — verbose test names
- `--tb=short` — concise tracebacks on failure
- `-x` — stop on first failure
- `--strict-markers` — reject unregistered markers
- `-ra` — show summary of all non-passing tests
- `--durations=5` — show the 5 slowest tests

## Environment Variables

| Variable | Default | Description |
|---|---|---|
| `TEST_DATABASE_URL` | `postgresql+asyncpg://postgres:postgres@localhost:5432/surfsense_test` | Database URL for tests |

The test suite forces `DATABASE_URL` to point at the test database, so your production database is never touched.

### Unit Tests

Pure logic tests that run without a database. Cover model validation, chunking, hashing, and summarization.

### Integration Tests

Require PostgreSQL + pgvector. Split into two suites:

- **`document_upload/`** — Tests the HTTP API through public endpoints: upload, multi-file, duplicate detection, auth, error handling, page limits, and file size limits. Uses an in-process FastAPI client with `ASGITransport`.
- **`indexing_pipeline/`** — Tests pipeline internals directly: `prepare_for_indexing`, `index()`, and `index_uploaded_file()` covering chunking, embedding, summarization, fallbacks, and error handling.

External boundaries (LLM, embedding, chunking, Redis) are mocked in both suites.

## How It Works

1. **Database setup** — `TEST_DATABASE_URL` defaults to `surfsense_test`. Tables and extensions (`vector`, `pg_trgm`) are created once per session and dropped after.
2. **Transaction isolation** — Each test runs inside a savepoint that rolls back, so tests don't affect each other.
3. **User creation** — Integration tests register a test user via `POST /auth/register` on first run, then log in for subsequent requests.
4. **Search space discovery** — Tests call `GET /api/v1/searchspaces` and use the first available space.
5. **Cleanup** — A session fixture purges stale documents before tests run. Per-test cleanup deletes documents via API, falling back to direct DB access for stuck records.

## Writing New Tests

1. Create a test file in the appropriate directory (`unit/` or `integration/`).
2. Add the marker at the top of the file:

```python
import pytest

pytestmark = pytest.mark.integration  # or pytest.mark.unit
```

3. Use fixtures from `conftest.py` — `client`, `headers`, `search_space_id`, and `cleanup_doc_ids` are available to integration tests. Unit tests get `make_connector_document` and sample ID fixtures.
4. Register any new markers in `pyproject.toml` under `markers`.