The right plan for every team.

Wrap your pipeline with server-side security scanning before any data is stored. Fluiq checks every prompt, response, tool call, and retrieved document, so attack patterns are never shipped in the public SDK.

• PII Detection & Redaction: Names, emails, phone numbers, SSNs, and credit cards, detected and redacted before persistence.
• Prompt Injection Blocking: Catches injection patterns in real time, before the prompt reaches your model.
• Jailbreak & Skeleton Key Detection: Dedicated scanners for jailbreak framings and skeleton-key attacks that try to unlock restricted behaviour.
• Semantic Attack Scoring: Similarity scoring against known attack vectors, so a reworded attack still scores as one.
• Secret Leak Prevention: Scans model output for leaked API keys, tokens, and high-entropy credential strings.
• Indirect Injection & RAG Poisoning: Inspects tool outputs and retrieved chunks for second-order injection, and flags documents that read like an attack.
• Tool Abuse Defense: Catches sensitive data exfiltrated through tool-call arguments, and tools invoked outside your allowlist.
• Multi-Agent Trust: Detects cross-agent injection and risk escalating across agent handoffs in the trace DAG.
• Image & Multimodal Scanning: Images and other media attached to a call are scanned alongside the text.
• Custom Guardrail Policies: Set your own thresholds and categories per organisation instead of taking the defaults.
• Warn or Block mode: warn (default) flags risks and attaches security metadata to the trace. block intercepts before the LLM call and raises FluiqSecurityError.

fluiq.instrument(api_key="fl_...")
fluiq.secure()  # warn mode flags risks on the trace
fluiq.secure(mode="block")  # block mode

Scoring runs server-side on traces Fluiq already has. Single responses get LLM-as-judge metrics; whole agent runs get judged on the decisions they made; golden datasets turn both into a regression gate.

• Six Judge Metrics: Hallucination, faithfulness, relevance, toxicity, coherence, and completeness, scored per response.
• Per-Metric Thresholds: Set a gate for each metric. warn logs the score on the trace; block raises FluiqEvalError before the response reaches your app.
• Agentic Evaluation: Judges a whole run in layers: deterministic checks, tool-selection quality, trajectory against the goal, and multi-agent coordination across the DAG.
• Multi-Model Jury: Borderline verdicts convene a panel of different judge models, with every member's score and reasoning kept for audit.
• Your Judge, Your Keys: Choose which model judges and which models sit on the panel, and bring your own provider key so judge tokens bill at the rate you negotiated.
• Dataset Regression Runs: Batch the same judges over a golden dataset of pinned trajectories, then diff one run against another to see exactly what got worse.
• Prompt Management: Version and deploy prompt templates, and promote any saved prompt to a custom judge by slug.
• Auditable Prompts: Every result carries the exact judge prompt and version that produced it, so a shifting metric traces back to a prompt change.

fluiq.instrument(api_key="fl_...")
fluiq.eval(thresholds={"hallucination": 0.8})  # warn mode
fluiq.eval(thresholds={"hallucination": 0.8}, mode="block")