Languages: English | 中文 | 日本語

llm-flow-dsl

AI + Business DSL as an AI rules engine.

AI + Business DSL for production workflows.

llm-flow-dsl is an AI rules engine: use a small, readable DSL to orchestrate LLM decisions, business constraints, approvals, retries, and tool calls in one place.

Why this can win

Most AI demos fail in production because teams cannot express policy, risk control, and business logic around LLM calls.

llm-flow-dsl focuses on the missing layer:

• Product teams can read and edit flow rules.
• Engineering teams can version, test, and release safely.
• AI behavior is governed by deterministic business rules.

Positioning

• Not "just another prompt framework".
• Not "just workflow automation".
• It is a domain language for AI decisions under business constraints.

One-line pitch:

Ship reliable AI features by writing business-grade decision flows as code.

Core design principles

• Readable first: PM/ops can understand every rule.
• Deterministic guardrails: model output is never the only source of truth.
• Production-native: retries, fallback models, approvals, audit log.
• Testable: every flow can run in CI with fixtures.

MVP scope (v0.1)

• DSL parser for flow blocks:
  • input
  • route
  • llm
  • if/else
  • tool
  • approval
  • output
• Local runner:
  • execute flow end-to-end
  • structured logs per step
  • dry-run mode
• Validation:
  • schema checks
  • unknown symbol checks
  • unsafe config checks
• Tests:
  • golden tests for flow outputs
  • deterministic fixtures for model/tool stubs

Killer first use-cases

• AI customer support triage with escalation policy
• AI sales lead qualification with compliance gates
• AI content moderation with region-specific rules
• AI internal copilot actions requiring approval workflow

Growth playbook (how to make it "hot")

1. Build one flagship demo that solves a painful business problem.
2. Open-source 10 practical flow templates by industry.
3. Publish "before/after" reliability metrics (not just flashy demos).
4. Launch a no-code visualizer for DSL execution traces.
5. Push weekly content: failure cases, guardrail patterns, production lessons.

30-day launch plan

1. Week 1: parser + runner + one real template.
2. Week 2: test harness + trace viewer CLI output.
3. Week 3: docs site + 3 industry examples.
4. Week 4: public launch + live teardown of real user flow.

Project structure

llm-flow-dsl/
  README.md
  docs/
    grammar-spec.md          (EBNF formal grammar)
    product-strategy.md      (business vision)
    roadmap.md               (development roadmap)
  examples/
    support-triage.flow      (customer support triage + escalation)
    ab-routing.flow          (A/B experiment routing with guardrails)
    refund-approval.flow     (refund triage with risk gating)
    policy-gating.flow       (region/content policy enforcement)
    content-moderation.flow  (multi-level severity moderation)
    lead-scoring.flow        (sales lead qualification + compliance)
    fraud-detection.flow     (transaction fraud risk assessment)
    draft-review.flow        (AI draft quality review pipeline)
    duplicate-symbols.flow   (validation error demonstration)
    data/
      *.json                 (input fixtures for dry-run)
  tests/
    test_mvp_parser.py       (parser + runner + validation tests)

Quick start

# Parse a flow file into AST (JSON)
python -m llm_flow_dsl parse examples/support-triage.flow --pretty

# Parse + validate (default: validation is on)
python -m llm_flow_dsl parse examples/content-moderation.flow --pretty

# Run a flow with dry-run mode (deterministic, no LLM needed)
python -m llm_flow_dsl run examples/support-triage.flow --input-json examples/data/support-triage-input.json --dry-run --pretty

# Override LLM output for deterministic testing
python -m llm_flow_dsl run examples/fraud-detection.flow --input-json examples/data/fraud-detection-override-input.json --dry-run --pretty

# Run tests
python -m pytest tests/test_mvp_parser.py -v

Grammar reference: docs/grammar-spec.md

Examples

Each .flow file demonstrates a real-world AI decision workflow with business constraints.

Customer Support Triage

examples/support-triage.flow Classify support tickets by intent and urgency. Routes to appropriate queues, escalates enterprise+high-urgency cases.
DSL features: input, llm, if/else, route, tool, approval, output, member access, and/or logic

A/B Routing

examples/ab-routing.flow Route users to different LLM variants based on experiment bucket. Demonstrates not in guardrail for unknown buckets.
DSL features: not in operator, route on expression, guardrail approval

Refund Approval

examples/refund-approval.flow Refund triage with fraud risk classification. Multi-tier approval based on amount and risk level.
DSL features: nested if/else, member access, boolean logic in output expressions

Policy Gating

examples/policy-gating.flow Enforce region-specific content policies. Route to different pipelines based on user region and content category.
DSL features: in/not in operators, route with tool/approval actions

Content Moderation

examples/content-moderation.flow Multi-level content severity routing (critical→approval, low→auto-approve). PII data protection for low-reputation authors.
DSL features: severity-based route, not in for safe categories, conditional approval chaining

Lead Scoring

examples/lead-scoring.flow Qualify sales leads with LLM scoring. Compliance gates for regulated industries (healthcare/finance).
DSL features: industry-based if/else, score-based route, compliance approval workflow, and conditions

Fraud Detection

examples/fraud-detection.flow Assess transaction fraud risk with multi-tier review. Triggers specialist approval for high-risk or large transactions.
DSL features: approval chains, not in for international routing, output expressions with or/and

Draft Review

examples/draft-review.flow Review AI-generated drafts for quality and safety. Fast-track publication for high-quality safe content.
DSL features: nested if/else, route on safety issues, guest author workflow, conditional fast-track

Validation Error Demo

examples/duplicate-symbols.flow Deliberately invalid flow demonstrating the validator's error detection (duplicate fields, duplicate LLM names).
Purpose: test and demonstrate semantic validation rules

Contributing

We welcome contributions! See CONTRIBUTING.md for:

• Development setup
• How to submit pull requests
• Code of Conduct

Early contributors will shape the DSL grammar and roadmap.

License

llm-flow-dsl is released under the MIT License. See LICENSE for details.

This means you can:

• ✅ Use it freely in commercial projects
• ✅ Modify and distribute the code
• ✅ Use it without asking permission

With the condition that you:

• ⚠️ Include the original copyright notice

Security

Found a security vulnerability? Please report it responsibly: 📧 See SECURITY.md for details.

Support

• 📖 Documentation
• 🐛 Report Issues
• 💬 Discussions

---

Built with ❤️ by the llm-flow-dsl community

中文简介

llm-flow-dsl 是一个面向生产环境的 AI 规则引擎。
你可以用简洁、可读的 DSL，把 LLM 决策、业务约束、审批、重试和工具调用统一编排在同一条流程里。

核心目标：

• 让产品、运营也能读懂并参与规则设计
• 让工程团队可以版本化、测试化、可审计地发布 AI 流程
• 用确定性的业务规则约束模型输出，提升稳定性与可控性

一句话：

用业务级决策流程代码，交付可靠的 AI 功能。

日本語概要

llm-flow-dsl は、本番運用向けの AI ルールエンジン です。
読みやすい小さな DSL で、LLM の判断、業務制約、承認、リトライ、ツール呼び出しを 1 つのフローとして記述できます。

主な狙い：

• プロダクト/オペレーション担当でもルールを理解・編集しやすい
• エンジニアがバージョン管理・テスト・監査可能な形で運用できる
• モデル出力だけに依存せず、決定論的な業務ルールで制御できる

一言で言うと：

ビジネス品質の意思決定フローをコード化し、信頼できる AI 機能を提供する。