obsidian/wiki/tech-patterns/cost-tracker-integration.md
2026-04-27 17:32:33 +01:00

300 lines
11 KiB
Markdown

---
title: AI Cost Tracker — Integrating a New Project
tags: [how-to, ai, cost-tracking, integration]
created: 2026-04-27
updated: 2026-04-27
updated3: 2026-04-27 (v3 — ContextVar pitfall, .env wipeout risk, docker inspect pattern)
---
# Integrating a New Project with AI Cost Tracker
Step-by-step guide for connecting any Oliver backend project to the shared cost-tracker.
## Service URLs
| Environment | Base URL |
|---|---|
| Dev | `https://optical-dev.oliver.solutions/cost-tracker/v1` |
| Prod (future) | `https://cost.oliver.agency/v1` |
Health check: `GET {base_url}/health``{"status":"ok","db":"ok"}`
## Prerequisites
- Access to `https://optical-dev.oliver.solutions/cost-tracker/` (Microsoft SSO or ask Vadym for dev access)
- Your project is a Python backend (FastAPI/other) — or adaptable for any HTTP client
---
## Step 1 — Get an API key
1. Open the Admin UI → **API Keys****Create key**
2. Name it after your project (e.g. `video-accessibility-dev`)
3. Copy the key — it is shown **only once**
4. Store in your project's `.env` as `COST_TRACKER_API_KEY`
---
## Step 2 — Add environment variables
```env
COST_TRACKER_BASE_URL=https://optical-dev.oliver.solutions/cost-tracker/v1
COST_TRACKER_API_KEY=ct_live_xxxxxxxxxxxxxxxxxxxx
COST_TRACKER_SOURCE_APP=your-project-name
```
> **Gotcha — Docker doesn't re-read `env_file` on `restart`.**
> After editing `.env`, run `docker compose up -d --force-recreate <service>`.
> Verify with:
> ```bash
> docker inspect <container> 2>/dev/null | python3 -c \
> "import sys,json; cfg=json.load(sys.stdin)[0]['Config']['Env']; [print(e) for e in cfg if 'COST' in e]"
> ```
> **Gotcha — leading spaces in `.env` = key silently ignored.**
> When appending via SSH, use `echo "KEY=val" >> .env` — never heredoc over SSH (shell quoting adds spaces).
> **Gotcha — `.env` can be silently wiped during `git pull` or Docker rebuild.**
> Projects that bake source into the image (`COPY . .` in Dockerfile, no volume mount) will overwrite the server's `.env` on any fresh build or pull.
> Always re-run `grep COST_TRACKER .env` AND the `docker inspect` command above after any redeploy.
---
## Step 3 — Create a lightweight HTTP client
No SDK package exists — use `httpx` directly. Create `core/cost_tracker.py`:
```python
import httpx
from app.core.config import settings
_client: httpx.AsyncClient | None = None
def get_client() -> httpx.AsyncClient:
global _client
if _client is None:
_client = httpx.AsyncClient(
base_url=settings.cost_tracker_base_url,
headers={"X-API-Key": settings.cost_tracker_api_key},
timeout=10.0,
)
return _client
```
Add to `Settings` in `core/config.py`:
```python
cost_tracker_base_url: str = "https://optical-dev.oliver.solutions/cost-tracker/v1"
cost_tracker_api_key: str = ""
cost_tracker_source_app: str = "my-project"
```
---
## Step 4 — Wrap AI calls (preflight → call → record)
This is the **core pattern**. Every paid AI call follows three steps.
> **Pitfall — do NOT use `contextvars.ContextVar` to pass user identity.**
> Python's `ContextVar` copies the context at task creation time. If your AI call is invoked via `asyncio.wait_for()`, `asyncio.create_task()`, or `asyncio.ensure_future()`, a `ContextVar` set in the calling coroutine will not be visible inside the spawned task — it silently returns the default value (`""` in most cases), so `user_external_id` ends up empty in the cost tracker.
> **Always pass `user_external_id: str` as an explicit function parameter** all the way down to the function that calls `record()`.
```python
from app.core.cost_tracker import get_client
import time
async def call_gemini_with_tracking(
prompt: str,
user_id: str,
job_id: str,
project_id: str | None = None,
) -> GenerateContentResponse:
ct = get_client()
# 1. Preflight — checks budget before making the AI call
preflight = await ct.post("/preflight", json={
"user_external_id": user_id,
"project_external_id": project_id,
"job_external_id": job_id,
"source_app": settings.cost_tracker_source_app,
"model": "gemini-2.5-pro-preview",
"estimated_units": {
"input_tokens": len(prompt) // 4, # rough estimate
"output_tokens": 2048,
},
})
preflight.raise_for_status()
pf = preflight.json()
if not pf["allow"]:
raise Exception(f"Budget exceeded: {pf.get('deny_reason')}")
# 2. Actual AI call
t0 = time.monotonic()
response = await client.models.generate_content(
model="gemini-2.5-pro-preview",
contents=prompt,
)
elapsed_ms = int((time.monotonic() - t0) * 1000)
# 3. Record actual usage
await ct.post("/usage/record", json={
"request_id": pf["request_id"],
"user_external_id": user_id,
"project_external_id": project_id,
"job_external_id": job_id,
"source_app": settings.cost_tracker_source_app,
"model": "gemini-2.5-pro-preview",
"units": {
"input_tokens": response.usage_metadata.prompt_token_count,
"output_tokens": response.usage_metadata.candidates_token_count,
},
"latency_ms": elapsed_ms,
"status": "success",
})
return response
```
For providers without token metadata (ElevenLabs, Google Cloud TTS), use `chars` instead:
```python
await ct.post("/usage/record", json={
...
"model": "eleven_multilingual_v2",
"units": {"chars": len(text)},
"latency_ms": elapsed_ms,
"status": "success",
})
```
See [[wiki/tech-patterns/cost-tracker-providers|cost-tracker-providers]] for all provider details.
---
## Step 5 — Handle errors gracefully (don't break the AI pipeline)
Cost tracking should never crash your main pipeline. Wrap calls:
```python
async def safe_preflight(ct, payload) -> dict:
"""Returns allow=True by default if cost-tracker is unreachable."""
try:
r = await ct.post("/preflight", json=payload)
r.raise_for_status()
return r.json()
except Exception as e:
logger.warning(f"Cost tracker preflight failed (allowing call): {e}")
return {"allow": True, "request_id": None, "estimated_cost_usd": 0}
async def safe_record(ct, payload) -> None:
try:
r = await ct.post("/usage/record", json=payload)
r.raise_for_status()
except Exception as e:
logger.warning(f"Cost tracker record failed (call already made): {e}")
```
> **Note:** For higher reliability, implement a SQLite outbox — save failed records to a local DB and retry on next request. This is optional for most projects.
---
## Step 6 — Sync users and projects (optional but recommended)
When a user is created or a project is configured in your app, mirror it to the cost-tracker so it appears in the analytics UI with a proper name:
```python
# On user create/login — field is "external_id" (NOT "user_external_id")
await ct.post("/users/upsert", json={
"external_id": str(user.id), # ← correct field name
"email": user.email,
"full_name": user.full_name,
# source_app is inferred from API key — do not send it
})
# On project create
await ct.post("/projects/upsert", json={
"external_id": str(project.id),
"source_app": settings.cost_tracker_source_app,
"name": project.name,
"workspace_id": "your-workspace-id", # from Admin UI
})
```
---
## Step 7 — Create Workspace / Team / Project in Admin UI
Before going live:
1. **Admin UI**`https://optical-dev.oliver.solutions/cost-tracker/`
2. **Workspaces** → Create (e.g. "Ford", "H&M", or "Oliver Internal")
3. **Teams** → Create under workspace (e.g. "Video Production")
4. **Projects** → Create under team, set `source_app` = your project name
5. **API Keys** → Create key, assign to the project
> Jobs sent before a project is created appear as **Unassigned** in the dashboard — assignable later in bulk.
---
## Step 8 — Set budgets and alerts
**Admin UI → Budgets → Create:**
- `scope_type`: workspace / team / project
- `amount_usd`: monthly limit
- `alert_thresholds`: `[0.5, 0.8, 1.0]` → email alerts at 50%, 80%, 100%
- `hard_limit`: `true` → preflight returns `allow=false` when exceeded
---
## Step 9 — Smoke test
```bash
BASE="https://optical-dev.oliver.solutions/cost-tracker/v1"
KEY="ct_live_your_key_here"
# 1. Health
curl "$BASE/health"
# → {"status":"ok","db":"ok"}
# 2. Preflight
curl -X POST "$BASE/preflight" \
-H "X-API-Key: $KEY" \
-H "Content-Type: application/json" \
-d '{"user_external_id":"test-user","source_app":"my-project","model":"gemini-2.5-pro-preview","estimated_units":{"input_tokens":1000,"output_tokens":200}}'
# → {"allow":true,"estimated_cost_usd":...,"request_id":"..."}
# 3. Record
curl -X POST "$BASE/usage/record" \
-H "X-API-Key: $KEY" \
-H "Content-Type: application/json" \
-d '{"request_id":"<from above>","user_external_id":"test-user","source_app":"my-project","model":"gemini-2.5-pro-preview","units":{"input_tokens":987,"output_tokens":180},"latency_ms":1200,"status":"success"}'
# → {"event_id":"...","cost_usd":0.00214}
```
Then open the Admin UI Dashboard — the test event should appear within seconds.
---
## Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| `401 Unauthorized` | Wrong or missing API key | Check `COST_TRACKER_API_KEY`; verify key active in Admin UI |
| `preflight.allow = false` | Budget exceeded | Admin UI → Budgets → raise limit |
| Events missing from dashboard | `source_app` mismatch | Ensure `source_app` matches in preflight, record, and project upsert |
| `cost_usd = null` on events | Model not in pricing table | Admin UI → Pricing → add model manually, or wait for LiteLLM sync |
| 502/504 from reverse proxy | cost-tracker containers down | `ssh optical-dev "docker compose -f /opt/ai-cost-tracker/infra/docker-compose.yml ps"` |
| Env vars not picked up after editing `.env` | `docker compose restart` doesn't re-read `env_file` | Use `docker compose up -d --force-recreate <service>` |
| Env vars in `.env` have leading spaces | Heredoc over SSH adds indentation | Use `echo "KEY=val" >> .env` per line; verify with `grep COST_TRACKER .env` |
| `source_app` shows in breakdown but "By Project" shows `—` | `project_id` null — not sent by integration | Use `source_app` dimension in Pivot Explorer instead of `project_id` |
| `user_external_id` is empty in all events | `ContextVar` not propagating across `asyncio.wait_for` / task boundary | Pass `user_external_id` as an explicit function parameter — do not use `ContextVar` |
| COST_TRACKER env vars disappear after deploy | `git pull` or Docker rebuild overwrites server `.env` | Re-add with `echo` commands; verify with `docker inspect` python3 one-liner after every rebuild |
## Related articles
- [[wiki/architecture/ai-cost-tracker|ai-cost-tracker architecture]] — full system diagram and design decisions
- [[wiki/tech-patterns/cost-tracker-pricing-sources|cost-tracker-pricing-sources]] — pricing pipeline
- [[wiki/tech-patterns/cost-tracker-providers|cost-tracker-providers]] — billing units per AI provider