Versioned templates with Jinja2 variables, A/B testing, AI recommendations, and per-key selection. Ship prompt changes without redeploying — and roll back in one click when they regress.
Active templates
Snapshot + diff + author + timestamp
nemo_prompt_template_id on any request
Hash-routed by request id
Honors org-level data policy
Template management with Jinja2 variables, version history, A/B testing, AI-powered recommendations, and a per-key selection model that works with the unmodified OpenAI SDK.
Author prompts with double-brace variables, conditionals, loops, filters, and default values. Inject runtime values via `nemo_prompt_variables` — never hardcode prompt text in your repo again.
Every save creates a versioned snapshot with author, timestamp, and diff. Activate any version instantly — no redeploy. If a change regresses output quality, roll back from the dashboard in one click.
Split traffic across two or more variants with a deterministic hash on request id. Each branch logs latency, cost, and response metadata so you can promote the winner with statistical confidence.
Get suggestions to clarify ambiguous instructions, cut token use, or fix common patterns that hurt model output. Recommendations are generated, never auto-applied — you stay in control of every change.
Apply any template at request time by passing `nemo_prompt_template_id` and `nemo_prompt_variables` in `extra_body`. No config flags, no redeploy — switch the prompt mid-request if you need to.
Who changed what, when, and from where. Every prompt save, activation, A/B start, and rollback is captured in the immutable audit log alongside the actor, IP, and full diff.
Don’t guess which version is better — measure it. Split traffic, capture the metrics that matter, and promote the winner. AI recommendations close the loop with predicted lift before you even ship.
A/B testing
Create two or more variants with a configurable traffic split. Routing is hash-deterministic per request id, so the same input always lands on the same variant. Per-branch metrics roll up in the analytics dashboard.
Variant performance (last 24h)
Closed-loop recommendations
Recommendations are generated, scored, and surfaced — but never auto-applied. You see the diff, the predicted impact, and the reasoning. Accept, edit, or dismiss; the audit trail captures every choice.
Open suggestions
Per-key + per-request
Pass `nemo_prompt_template_id` and `nemo_prompt_variables` in `extra_body` on any chat completion. Nemo Backend reads the template, renders with your variables, and forwards a clean payload to the provider — the prompt fields never leave the gateway.
Per-request override
Pulling prompts out of the codebase and into a versioned template store is the single biggest velocity unlock we have shipped this year. PMs change copy now; engineering ships features.
Engineering Manager
Mid-market SaaS, agent-platform team
Templates, versions, and variable payloads are stored as org-scoped rows in nemo.prompt_templates and nemo.prompt_versions, both protected by Row-Level Security. Cross-org reads are impossible at the database layer.
Tenant isolation
The same UUID flows through every layer. Prompt rows reference organization_id; RLS policies on every table refuse cross-tenant reads. The application can have bugs; the database refuses to return rows from the wrong tenant.
Prompt-related tables
Ship without redeploying
Free to start. Versioning, variants, recommendations, and the audit trail are unlocked on every plan from day one.