Prompts

1. Discover — the Prompts page surfaces every LLM call from your traces as a row. Click any row to open it in the evaluation playground alongside its full trace tree.
2. Edit — refine the template in the editor. Add {{variable}} placeholders; the UI detects them and renders input fields for test values.
3. Evaluate — run LLM-as-judge metrics (hallucination, faithfulness, relevance …) on the template + response pair. Results appear inline as scored cards.
4. Save — give the prompt a name and a unique slug. Every subsequent edit creates a version snapshot so you can restore any previous state.
5. Promote — deploy to development, staging, and production independently. Each environment stores a full snapshot of the template at promote time, so rolling back is one click.

Wrap any dynamic value in double curly braces. Variable names must start with a letter or underscore and contain only alphanumeric characters and underscores.

# Prompt template stored in the dashboard:
You are a helpful assistant for {{company}}.
Answer the following question in {{language}}: {{question}}

The playground detects variables automatically and renders a labeled input for each one so you can test substitutions before promoting.

Each named environment stores an independent snapshot — promoting to staging never touches production. The typical promotion flow:

Call fluiq.fetch_prompt() anywhere in your application to retrieve the deployed template for an environment. The call is authenticated with your API key and returns the snapshot that was promoted — not the current editor draft.

import fluiq

fluiq.instrument(api_key="fl_...")

# Fetch the production snapshot (default):
prompt = fluiq.fetch_prompt("customer-support-reply")

# Fetch a specific environment:
prompt = fluiq.fetch_prompt("customer-support-reply", env="staging")

# Fill template variables and call your LLM:
filled = prompt.render(
    company="Acme Corp",
    language="French",
    question=user_input,
)
response = client.chat.completions.create(
    model=prompt.model or "gpt-4o",
    messages=[{"role": "user", "content": filled}],
)

The returned object exposes:

Every time you save an edited template, the previous version is automatically snapshotted. Open the History panel on any saved prompt to browse past versions — each shows its version number, the template preview, the model, and when it was saved. Click Restore to roll back; the current state is snapshotted first so no work is ever lost.

Environments pin to their snapshot independently — restoring v3 to the head does not change what production is serving until you explicitly re-promote.