Regressa logs every LLM call, versions every prompt change, scores output quality automatically, and alerts you the moment quality drops, cost spikes or latency degrades.
Every distinct template hash becomes a version. Compare quality, latency, error rate and cost across versions, with a line diff.
LLM-as-judge with rubrics, semantic similarity to a golden set, or your own sandboxed scoring function. Sampled per template.
Thresholds, percent change, rolling-baseline anomaly detection and budget caps. Slack, signed webhooks, email.
npx @regressa/node gate blocks a merge when a prompt change regresses against production. Ships as a GitHub Action.
PII redaction at ingest, per-project retention, budget-driven sampling, OIDC SSO, roles and a full audit log.
SDKs for Node, Python, Go and Java that wrap OpenAI and Anthropic clients in two lines, plus native OpenTelemetry ingestion.
Wrap the client you already use. Declare the template and Regressa does the rest.
import OpenAI from "openai";
import { Regressa, wrapOpenAI } from "@regressa/node";
const regressa = new Regressa(); // REGRESSA_API_KEY
const openai = wrapOpenAI(new OpenAI(), regressa);
await openai.chat.completions.create({
model: "gpt-4o-mini",
messages,
regressa: { promptTemplate: { name: "support-agent", raw: TEMPLATE } },
});from openai import OpenAI
from regressa import Regressa, wrap_openai
regressa = Regressa() # REGRESSA_API_KEY
openai = wrap_openai(OpenAI(), regressa)
openai.chat.completions.create(
model="gpt-4o-mini",
messages=messages,
regressa_prompt_template={"name": "support-agent", "raw": TEMPLATE},
)client := regressa.New(regressa.Options{}) // REGRESSA_API_KEY
defer client.Shutdown(ctx)
span := client.Start("gpt-4o-mini", regressa.OpenAI, msgs,
regressa.WithTemplate("support-agent", template))
resp, err := provider.Call(ctx, req)
span.End(resp.Text, resp.PromptTokens, resp.CompletionTokens, err)Regressa regressa = Regressa.builder().build(); // REGRESSA_API_KEY
Regressa.Span span = regressa.start("gpt-4o-mini", "openai")
.message("system", prompt).message("user", question)
.template("support-agent", TEMPLATE);
try { span.end(out, promptTokens, completionTokens); }
catch (Exception e) { span.fail(e); throw e; }Stateless ingest, queued workers, TimescaleDB for time series, Next.js dashboard.
git clone https://github.com/dwarka-prasad/regressa && cd regressa
cp .env.example .env && pnpm install
pnpm infra:up && pnpm db:migrate && pnpm db:seed
pnpm dev # dashboard :3100 · ingest :4100 · workerSelf-host in minutes. Gate every prompt change in CI.