promptpolygraph 0.6.7

Synthetic-prompt evaluation and persona-audit harness for web and API AI systems.

PromptPolygraph

Synthetic-prompt evaluation and persona-audit harness for web and API AI systems.

PromptPolygraph pushes thousands of synthetic prompts through any web/API/LLM system, scores the responses against a pluggable rubric (plus cheap deterministic assertions and an optional multi-judge ensemble), reacts to them through a panel of personas, traces low scores with a forensic audit, and renders the result as a markdown / docx / pdf / html report.

It is local-first and cloud-agnostic: one process, an async concurrency pool, and a SQLite store — runs on a laptop, a CI runner, or a container. No work queue, no database server, nothing to deploy. It also ships a deployable API + worker (Postgres-backed, Dockerized) for multi-user, team-scale use — see Service mode.

Status: the local harness (CLI) and the service layer are both built and tested. Runs fully offline in --mock mode; the Docker image and the postgres + api + worker compose stack are verified end to end.

Why

Evaluating an AI system well means more than a single accuracy number. PromptPolygraph combines:

• Synthetic corpora — a fixed deterministic set for clean run-over-run baselining, or LLM-generated varied / adversarial probes, with dials for quantity, categories, difficulty, and seed.
• Adapters — one small class (async def query(case) -> Response) per system under test. Ships an HTTP/REST adapter, an LLM-chat adapter, and an in-process callable adapter.
• Layered scoring — deterministic assertions (contains / regex / json-schema / latency / semantic similarity / custom Python), with weights, per-test thresholds, and named + derived (F1-style) metrics, then an LLM rubric scorer with a multi-judge ensemble + inter-judge agreement.
• Persona panel — a pool of distinct individuals react to the real responses (trust, usefulness, clarity, would-return), reconciled against the rubric so you optimize real value, not just the number.
• Forensic audit — per-category agents trace low scores to failure modes and emit a concrete suggested fix (the file:line to change + a diff) for systems whose source you provide.
• Comparison & trends — comparability-gated N-run comparison, per-dimension trends, and regression detection vs a pinned or rolling baseline.
• Visuals & reports — a local dashboard (score heatmaps, compare matrix, trend lines, persona radar, root-cause→fix view) and presentation-grade markdown / docx / pdf / html reports with inline charts.

See CHANGELOG.md for what's new in each release.

Install

pip install -e ".[dev]"            # core + test deps
pip install -e ".[dev,llm]"        # + OpenAI-compatible adapter
pip install -e ".[dev,redteam]"    # + OSS-grounded red-team sources (garak, PyRIT, DeepTeam, datasets)

Requires Python 3.10+. The red-team engine and Arena work with no extras (the built-in catalog probe source needs no dependencies); the [redteam] extra only lights up the external OSS sources.

Quickstart

# Offline end-to-end demo against the bundled example (no API key needed):
# runs a fixed corpus through a built-in demo target, scores it, runs the persona
# + forensic audit, and writes md/html/docx/pdf reports under polygraph_out/.
polygraph all --config examples/everyday_assistant/config.yaml --mock --format md,html,docx,pdf

With an ANTHROPIC_API_KEY set and a real adapter configured, drop --mock.

Bundled example packs (each a self-contained config + corpus + rubric + personas):

• examples/everyday_assistant/ — the default: a neutral general-purpose assistant (general knowledge, how-tos, reasoning, refusal, safety, edge input).
• examples/support_bot/ — a customer-support assistant for a fictional SaaS.
• examples/clinical_trials/ — a domain pack for a clinical-trial-protocol assistant (personas for medical writer, clinical scientist, data manager, medical director, regulatory, biostatistician, clinical ops, pharmacovigilance).

Copy one as a starting point, or generate a fresh tailored pack with polygraph tune (see below).

Runbook

1. Point an adapter at your system

The adapter is the only thing that changes per target. Set it in config.yaml:

# HTTP / REST
adapter:
  type: http
  options:
    url: https://api.example.com/chat
    method: POST
    headers: { Authorization: "Bearer ${API_KEY}" }   # ${VAR} reads the env
    body_template: { message: "{{prompt}}" }           # {{prompt}} {{category}} {{id}}
    response_path: "reply.text"                         # JMESPath into the JSON response

# LLM chat (OpenAI / Anthropic / OpenAI-compatible; openai needs the [llm] extra)
adapter:
  type: llm
  options: { provider: anthropic, model: claude-opus-4-8, system: "You are ...", max_tokens: 512 }

For anything exotic, implement async def query(self, case) -> Response and pass an in-process callable with --callable mymodule:my_fn.

Run on local / open models (Ollama, vLLM, LM Studio)

The grader/judge, audit agents, and corpus generation can run against any OpenAI-compatible endpoint — not just Anthropic — so the whole eval can run 100% locally on open models, no API key:

llm:
  provider: ollama          # anthropic (default) | openai | ollama | vllm | lmstudio | openai-compatible
  base_url: http://localhost:11434/v1   # default for ollama; set for other servers
model: llama3.1             # the model NAME for the judge/audit/generation

A local provider (ollama, …) needs no key and runs live; Anthropic/OpenAI fall back to --mock when their key is absent. To also point the system under test at a local model: adapter: { type: llm, options: { provider: ollama, model: llama3.1 } }.

2. Choose a corpus mode (dials)

polygraph run --config c.yaml --mode fixed                       # deterministic set; stable ids for baselining
polygraph run --config c.yaml --mode varied      --count 1000    # fresh LLM-generated each run
polygraph run --config c.yaml --mode adversarial --count 500 --difficulty aggressive
polygraph run --config c.yaml --mode hybrid       --count 800    # fixed core + generated supplement

Dials: --count, --per-category, --categories a,b,c, --difficulty {mild|standard|aggressive}, --concurrency, plus rps/timeout/retries/resume in the config. Thousands of prompts is just the concurrency pool; runs are durable and --resume-able.

3. Score

polygraph analyze --config c.yaml --run <run_id> --judges 3 --ci

Deterministic assertions (declared per probe: contains / regex / json-schema / latency …) run first; then an LLM rubric scorer. --judges N adds an ensemble with inter-judge agreement. --ci exits non-zero if the gate fails (per-dimension threshold, cumulative across categories). The rubric is a YAML pack — edit rubric.yaml (dimensions, anchors, per-category applicability, threshold) for your domain.

4. Audit (persona panel + forensic)

polygraph audit --config c.yaml --run <run_id>

A panel of personas reacts to the real responses (trust / usefulness / clarity / would-return) and is reconciled against the rubric so you optimize real value, not just the number; per-category forensic agents trace low scores to the highest-leverage fixes.

Manage personas:

polygraph personas list
polygraph personas new "a grumpy retiree who hates phone trees" --out persona.yaml
polygraph personas generate --domain "developer tools" --count 8 --out panel.yaml

Select a panel via personas_path: (a YAML file), audit.persona_pool: N (sample the 13-persona library), or audit.personas: [ids].

5. Report, compare, trend & regressions

polygraph report  --config c.yaml --run <run_id> --format md,docx,pdf,html --baseline <prior_run_id>
polygraph compare --config c.yaml --run-a <id> --run-b <id>        # A/B win/loss/tie
polygraph compare --config c.yaml --runs <id1>,<id2>,<id3>         # N-run comparison report
polygraph trend   --config c.yaml --project <name> --window 30     # per-dimension trend over runs
polygraph regressions --config c.yaml --run <id> --against rolling:20   # or --against <baseline_id>

--baseline adds Δ-vs-prior columns + regression flags. Comparison is comparability-gated (runs are only score-compared when their corpus + rubric fingerprints match). trend plots per-dimension means + slope over the recent comparable runs; regressions flags dimensions that dropped vs a pinned or rolling baseline. PDF uses LibreOffice headless when available and is skipped gracefully otherwise.

One-shot

polygraph all --config c.yaml --format md,html        # run -> analyze -> audit -> report

Browse results in a local dashboard

A zero-dependency web UI (stdlib only — no server to deploy) over the runs the CLI has produced:

polygraph dashboard --out-dir polygraph_out          # opens http://127.0.0.1:8765

Browse runs, drill into category/dimension scores, per-case responses + assertions, the persona panel and forensic audit, and open the rendered reports. (For a multi-user, Postgres-backed team server with a job queue + API, see Service mode.)

Export the generated prompts as a reusable dataset

PromptPolygraph generates synthetic corpora — export them as a portable dataset to reuse anywhere:

polygraph export --run <run_id> --out prompts.json                 # full cases from a stored run
polygraph export --corpus mycorpus/ --out prompts.jsonl --format jsonl
polygraph export --corpus mycorpus/ --out prompts.csv --format csv --prompts-only

Formats: json / jsonl / csv; --prompts-only emits just {prompt, category} for a clean prompt corpus.

Customize reports (templates & branding)

Reports render from Jinja2 templates you can override and brand:

report:
  template: default            # built-in: default | minimal
  template_dir: my_templates/  # optional: your own report.html.j2 / report.md.j2 (overrides built-in)
  branding: { title: "Acme Eval", accent: "#0aa", logo: "https://…/logo.png" }

Tailor it to your domain

Set a domain (a one-line description of the system under test) and the generated prompts — and, with audit.tailor_personas, the persona panel — are specific to it instead of generic:

polygraph all --config c.yaml --mode adversarial --count 500 \
  --domain "Clinical trial protocol digitalization assistant"

Or scaffold a whole tailored project (rubric + persona panel + starter corpus + config) for a domain in one command, then refine the files and point the adapter at your system:

polygraph tune --domain "Clinical trial protocol digitalization assistant" --out projects/ctpd
polygraph all  --config projects/ctpd/config.yaml --mock

Bootstrap a golden set with a subject-matter expert

polygraph tune auto-scaffolds; polygraph elicit builds an expert-validated golden corpus through a guided, human-in-the-loop walkthrough:

polygraph elicit interview --domain "Clinical trial protocol digitalization assistant" --out brief.yaml
#   (or: polygraph elicit init --domain "..." --out brief.yaml   — fill the brief by hand/async)
polygraph elicit build    --brief brief.yaml --out projects/ctpd     # draft probes + review sheet + rubric + personas
#   the SME edits projects/ctpd/review.yaml (set `decision: reject` to drop a probe, or edit any field)
polygraph elicit finalize --review projects/ctpd/review.yaml --out projects/ctpd   # accepted probes -> golden corpus
polygraph all --config projects/ctpd/config.yaml --mock

The interview asks the expert what the system does, what good/bad answers look like per category, the failure modes and must-refuse cases, and real example queries; probes are drafted grounded in those answers, then gated by the expert's review — which is what makes the set golden. Personas are drawn from the expert roles named in the brief.

A ready-made clinical-trials pack ships under examples/clinical_trials/ (personas for medical writer, clinical scientist, data manager, medical director, regulatory, biostatistician, clinical ops, and pharmacovigilance; a protocol-focused corpus and rubric):

polygraph all --config examples/clinical_trials/config.yaml --mock --format md,html

Red-team & Arena

An authorized adversarial red team for a system you own — a roster of attacker agents pressurizes the target, a breach judge rules each exchange, and findings roll up into a severity-ranked vulnerability report mapped to the OWASP Top-10 for LLM Apps and MITRE ATLAS.

# Offline demo against the built-in target (zero tokens, deterministic):
polygraph redteam --profile all_frontier --mock --format md,html,json

# Ground the attack in OSS tooling + datasets, judge with a safety classifier:
polygraph redteam --profile deep \
  --sources catalog,garak,pyrit,deepteam,dataset:advbench \
  --guard --format md,html,json

polygraph redteam profiles      # list the preconfigured teams

• Preconfigured teams — all_frontier (default; every strategy on a frontier model, each agent a distinct persona + temperature), multi_frontier (mixed vendors), mixed (local open + frontier), local_swarm (100% offline), pressure, quick, and deep (adaptive smart multi-turn).
• OSS sources — attackers seed + mutate from a catalog of known techniques. --sources folds in optional garak, PyRIT, DeepTeam, and on-demand datasets (AdvBench / HarmBench / JailbreakBench); each runs through the same target → judge loop. Install the [redteam] extra.
• Smart multi-turn — PAIR iterative refinement, Crescendo escalation, and TAP candidate trees, per-attacker mode.
• Converters — base64 / rot13 / leetspeak / payload-split / roleplay-wrap / many-shot transforms, per-attacker converter.
• Judges — an LLM reviewer (default) or --guard for a Llama-Guard-style safety classifier on a local model.
• Arena — the dashboard (/redteam, SSE) or service (/ws/redteam, WebSocket) streams the run live, persists it for replay (Live/Replay toggle), and offers a per-attacker drill-down: the multi-turn timeline plus a root cause that — when redteam.code_path points at a local checkout and a model is available — traces the breach to file:line rungs of the target source (local model by default; air-gap and consent controls for any remote provider). Without a clone it shows the finding summary (control, OWASP/ATLAS, mitigation).

The report leads with the attack success rate (ASR), an OWASP coverage table (tested vs breached), and OWASP/ATLAS tags on every finding. See docs/REDTEAM_LOCAL.md for running attacker/judge models locally (Ollama / HuggingFace, Mac Studio / NVIDIA DGX Spark / cloud).

Architecture

One local process: Corpus → Runner (async pool, timeout/retry/resume) → Adapter → target → Response → SQLite + cache → Analyze (assertions + LLM ensemble) → Gate → Audit (persona + forensic) → Report. No work queue, no database server. The store and dispatch sit behind interfaces, so a service/worker deployment can wrap the same engine without touching it.

The red-team engine reuses the same adapter + report spine: Profile (attacker roster) → [LLM attackers ⨯ multi-turn modes] + [OSS sources] → Adapter → target → Breach judge (LLM reviewer | Llama-Guard) → Vulnerability report (ASR + OWASP/ATLAS), with a live event stream (emit) that the CLI logs and the dashboard/service render as the Arena.

Service mode (multi-user, scalable)

The same engine runs as a deployable API + worker for a team:

pip install -e ".[service]"
polygraph-server      # FastAPI: trigger runs, poll status, fetch reports, compare, manage personas
polygraph-worker      # claims queued jobs and executes the pipeline (scale horizontally)
# or one container that does both (in-process worker) for local use

Or the whole stack in containers (Postgres + API + a separate worker):

docker compose up --build           # api on :8080, worker scales with --scale worker=N
curl -XPOST localhost:8080/api/runs -H 'X-API-Key: change-me-dev-key' \
     -H 'Content-Type: application/json' -d '{"config_name":"support_bot"}'

The default Docker image includes LibreOffice, so PDF reports render in-container; build with --build-arg INCLUDE_PDF=false for a slimmer image (html/md/docx still work).

• Store: SQLite locally, Postgres in production (same code, just POLYGRAPH_DATABASE_URL). A durable job queue lets many workers pull work safely (FOR UPDATE SKIP LOCKED on Postgres).
• API (/api, API-key auth): create runs, poll status + live progress, fetch summary/report (html & markdown re-render from the DB), browse cases, A/B compare, manage personas. Plus a dashboard at /.
• Scheduling & CI: cron schedules enqueue recurring runs; per-run or global webhooks POST a summary on completion; --ci-style gating is available via the run verdict.
• Deploy: one Docker image, two roles, on AWS (App Runner / ECS Fargate / EKS) or GCP (Cloud Run).

See docs/SERVICE.md for the operator runbook and deploy/README.md for cloud deployment.

Development

pip install -e ".[dev,service]"
pytest -q                                   # full suite
python -m promptpolygraph --help            # same as the `polygraph` CLI
python -m build                             # build wheel + sdist

CI (.github/workflows/ci.yml) runs the suite on Python 3.11/3.12, smoke-tests the default example offline, and verifies the built wheel ships the persona data.

License

Apache-2.0. See LICENSE.