feat: add run.sh to f1_radio_rag (#19)

## Summary

Adds CI support to `use-cases/f1_radio_rag` — the first example to go
through the new PR test workflow (from PR #16).

## What changed

- `run.sh` — runs `ingest` then `ask`, logging one RAG trace to Opik.
Lightweight alternative to the full eval suite for CI.
- `config.py` — reads `OPIK_EXAMPLES_MODEL` so CI routes to a cheap
model (`openai/gpt-4o-mini`). Removes `LLM_READY` guard — missing keys
now fail loudly via litellm rather than silently skipping.
- `cli.py` — removes the `LLM_READY` dry-run branch in `ask`.

## Tested locally

Ran `bash run.sh` with real credentials. Trace appeared in Opik
workspace under project `f1-radio-rag`.

**Depends on PR #16 merging first** for the PR test workflow to fire.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Author: LeoRoccoBreedt

templates/use-case-template/README.md

@@ -33,7 +33,7 @@ Or, with `uv` (recommended — this folder is a `uv` project): `uv sync`.
 
 | Environment variable | Required | Description |
 |---|---|---|
-| `ANTHROPIC_API_KEY` | for `run`/`eval`/`optimize`/`promote` | Anthropic key; used via litellm for generation, judge metrics, and the optimizer |
+| `ANTHROPIC_API_KEY` (or the key for your `OPIK_EXAMPLES_MODEL` provider) | for live `run`/`eval`/`optimize`/`promote` | Model-provider key used via litellm for generation, judge metrics, and the optimizer |
 | `OPIK_API_KEY` | for `eval`/`optimize`/`promote` | Your Opik API key. Unset → those commands run in DRY_RUN |
 | `OPIK_WORKSPACE` | for `eval`/`optimize`/`promote` | Your Opik workspace name |
 | `OPIK_PROJECT_NAME` | No | Opik project for traces/experiments (default `example-use-case`) |

templates/use-case-template/src/example_use_case/cli.py

@@ -8,9 +8,9 @@
 
 @app.command()
 def run(input: str, context: list[str] = typer.Option(None, "--context", "-c")) -> None:
-    """Run the app on a single input (traced in Opik when an LLM key is set)."""
-    if not config.LLM_READY:
-        typer.echo("[DRY RUN] ANTHROPIC_API_KEY not set — would run the app on:")
+    """Run the app on a single input (traced in Opik when credentials are set)."""
+    if config.DRY_RUN:
+        typer.echo("[DRY RUN] Opik creds not set — would run the app on:")
         typer.echo(f"  input: {input}")
         for line in context or []:
             typer.echo(f"  context: {line}")

templates/use-case-template/src/example_use_case/config.py

@@ -3,12 +3,9 @@
 OPIK_API_KEY = os.environ.get("OPIK_API_KEY")
 OPIK_WORKSPACE = os.environ.get("OPIK_WORKSPACE")
 OPIK_PROJECT_NAME = os.environ.get("OPIK_PROJECT_NAME", "example-use-case")
-ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY")
 
 # No Opik credentials -> every command still runs and prints locally instead of calling Opik.
 DRY_RUN = not (OPIK_API_KEY and OPIK_WORKSPACE)
-# Generation/judging needs an Anthropic key; without it we describe what would happen.
-LLM_READY = bool(ANTHROPIC_API_KEY)
 
 # litellm model strings (Anthropic provider). Swap for any litellm-supported model.
 # CI sets OPIK_EXAMPLES_MODEL to a cheap model; locally, leave it unset to use the full model.

use-cases/f1_radio_rag/README.md

@@ -38,10 +38,11 @@ Or, with `uv` (recommended — this folder is a `uv` project): `uv sync`.
 
 | Environment variable | Required | Description |
 |---|---|---|
-| `ANTHROPIC_API_KEY` | for `ask`/`eval`/`optimize`/`promote` | Anthropic key; used via litellm for generation, the LLM-judge metrics, and the optimizer |
+| `ANTHROPIC_API_KEY` (or the key for your `OPIK_EXAMPLES_MODEL` provider) | for live `ask`/`eval`/`optimize`/`promote` | Model-provider key used via litellm for generation, the LLM-judge metrics, and the optimizer |
 | `OPIK_API_KEY` | for `eval`/`optimize`/`promote` | Your Opik API key. Unset → those commands run in DRY_RUN |
 | `OPIK_WORKSPACE` | for `eval`/`optimize`/`promote` | Your Opik workspace name |
 | `OPIK_PROJECT_NAME` | No | Opik project for traces/experiments (default `f1-radio-rag`) |
+| `OPIK_EXAMPLES_MODEL` | No | litellm model for generation/judging/optimising. Unset → `anthropic/claude-sonnet-4-6`; CI sets a cheap model (e.g. `openai/gpt-4o-mini`) |
 | `OPIK_URL_OVERRIDE` | No | Base URL for self-hosted Opik (default: Opik Cloud) |
 
 ## Running it
@@ -66,13 +67,22 @@ uv run f1rag promote     # optimised prompt saved to the Prompt Library (version
 uv run f1rag run-all     # the whole loop in one shot
 ```
 
+`run.sh` is the entrypoint CI runs: it exports `OPIK_PROJECT_NAME`, then `uv sync` and runs
+`ingest` + `ask`. With no credentials it stays in dry-run and exits 0 (the secrets-free CI check);
+with credentials set it logs a live trace to Opik.
+
+```bash
+bash run.sh
+```
+
 ## How it works
 
 1. **Ingest** (`rag.py`) — `chromadb.PersistentClient` stores one document per radio message with
    session/driver/lap metadata, using ChromaDB's default local embeddings (no embedding-API cost).
 2. **Ask** (`rag.py`) — `answer()` retrieves the top-k messages, then calls
-   `litellm.completion(model="anthropic/claude-sonnet-4-6", ...)` with the summariser prompt from
-   `prompts.py`. It's decorated with `@opik.track`, so each call appears as a trace in Opik.
+   `litellm.completion(model=config.GEN_MODEL, ...)` with the summariser prompt from `prompts.py`
+   (`GEN_MODEL` defaults to `anthropic/claude-sonnet-4-6`, overridable via `OPIK_EXAMPLES_MODEL`).
+   It's decorated with `@opik.track`, so each call appears as a trace in Opik.
 3. **Eval** (`evaluation.py`) — builds an Opik dataset and a test suite, then scores the live RAG
    task. The test suite checks plain-English **assertions**; `evaluate` runs the `ContextRecall`
    (retrieval quality) and `Hallucination` (faithfulness) metrics. The eval cases live in

use-cases/f1_radio_rag/run.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+set -e
+
+export OPIK_PROJECT_NAME="f1-radio-rag"
+
+uv sync
+
+uv run f1rag ingest
+uv run f1rag ask "Why did Verstappen pit early?"

use-cases/f1_radio_rag/src/f1_radio_rag/cli.py

@@ -23,8 +23,8 @@ def ask(query: str, k: int = 5) -> None:
     typer.echo("Retrieved messages:")
     for message in context:
         typer.echo(f"  - {message}")
-    if not config.LLM_READY:
-        typer.echo("\n[DRY RUN] ANTHROPIC_API_KEY not set — would summarise the above with Claude.")
+    if config.DRY_RUN:
+        typer.echo(f"\n[DRY RUN] Opik creds not set — would summarise with {config.GEN_MODEL}.")
         return
     result = rag.answer(query, k)
     typer.echo(f"\nSummary:\n{result['output']}")

use-cases/f1_radio_rag/src/f1_radio_rag/config.py

@@ -3,17 +3,15 @@
 OPIK_API_KEY = os.environ.get("OPIK_API_KEY")
 OPIK_WORKSPACE = os.environ.get("OPIK_WORKSPACE")
 OPIK_PROJECT_NAME = os.environ.get("OPIK_PROJECT_NAME", "f1-radio-rag")
-ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY")
 
 # No Opik credentials -> every command still runs and prints locally instead of calling Opik.
 DRY_RUN = not (OPIK_API_KEY and OPIK_WORKSPACE)
-# Generation/judging needs an Anthropic key; without it we describe what would happen.
-LLM_READY = bool(ANTHROPIC_API_KEY)
 
-# litellm model strings (Anthropic provider).
-GEN_MODEL = "anthropic/claude-sonnet-4-6"  # deployed summariser
-JUDGE_MODEL = "anthropic/claude-sonnet-4-6"  # LLM-as-judge for metrics
-OPTIMIZER_MODEL = "anthropic/claude-sonnet-4-6"  # meta-model that rewrites the prompt
+# litellm model strings. CI sets OPIK_EXAMPLES_MODEL to a cheap model (e.g. openai/gpt-4o-mini);
+# locally, leave it unset to use the full model.
+GEN_MODEL = os.environ.get("OPIK_EXAMPLES_MODEL", "anthropic/claude-sonnet-4-6")  # deployed summariser
+JUDGE_MODEL = GEN_MODEL  # LLM-as-judge for metrics
+OPTIMIZER_MODEL = GEN_MODEL  # meta-model that rewrites the prompt
 
 CHROMA_DIR = "chroma_db"
 COLLECTION = "f1_radio"