⚖️ PatentFlow — Offline Agentic Patent Prosecution Workspace

PatentFlow is an enterprise-grade, privacy-first Document Processing Workspace designed for European Patent Attorneys. It targets the realities of prosecution work:

• Art. 123(2) EPC risk (added matter) where wording choices can be fatal
• Art. 56 EPC inventive-step mapping where semantic interpretation matters
• Client confidentiality where “cloud by default” is not acceptable

Built by an IP professional, for IP professionals.

---

Why PatentFlow

1) Legal accuracy under institutional constraints

Patent prosecution is not “generic writing.” It is risk management:

• A single phrasing shift can trigger an Art. 123(2) issue
• Inventive-step reasoning requires structured, repeatable mapping
• Quality and traceability matter more than “chatty” UX

2) 100% offline operation for client confidentiality

PatentFlow is designed to run fully locally:

• Local LLM execution (air-gapped capable)
• No external SaaS dependencies required for core workflows
• Local persistence for attorney-specific preferences

3) Enterprise UX: minimal, information-dense, institutional

The UI follows a Bloomberg Terminal-style aesthetic:

• High signal density
• Subtle controls
• Low-friction review of structured outputs

---

Trade Secret / Black Box Disclaimer (Intentional)

Specific system prompts, proprietary dictionaries, and heuristic parsing algorithms are intentionally omitted from this public repository to protect intellectual property.

PatentFlow exposes stable interfaces and deterministic boundaries while keeping core prompt logic and proprietary linguistic assets internal.

---

System Architecture (High-Level)

graph TD
    subgraph Frontend [Next.js Enterprise UI - Port 3000]
        UI[Workspace Dashboard]
        UI -->|POST /api/generate| API[FastAPI Gateway :8000]
        UI -->|GET /api/status/:id| API
        UI -->|GET/POST /api/memory/*| API
    end

    subgraph Backend [FastAPI + Celery Workers]
        API -->|Enqueue Tasks| Broker[(Redis Broker :6379)]
        API -->|Fetch Results| BackendRedis[(Redis Result Backend)]
        Broker -->|Consume| Worker[Celery Worker]

        Worker --> Skills[Skills Interface]
        SQLite[(Local Profile DB)] --> Skills
    end

    subgraph External [Optional Data Sources]
        EPO[EPO API] -->|Prior Art Retrieval| API
    end

    subgraph AI [Local AI Runtime]
        Skills --> LLM[(Local LLM)]
    end

Loading

---

Core Capabilities

1) Art. 56 Claim Chart Generation (LLM-Assisted, Structured Output)

Generate an attorney-reviewable claim chart with:

• Feature-by-feature claim splitting
• Prior art excerpt anchoring (D1/D2)
• LLM semantic assessment:
  • Yes / No / Partial
  • reasoning captured per row for auditability

2) Art. 123(2) Translation Verification (High-Risk Terminology Guardrails)

A verification workflow designed to surface:

• semantic mismatches
• risky wording drift
• institutional terminology consistency

3) Dynamic Attorney Memory (Local Persistent Context Injection)

PatentFlow supports a Local User Preference Engine that stores and recalls attorney preferences entirely offline:

• SQLite-backed memory (zero external dependencies)
• Profile-specific preferences persisted across sessions
• Preferences are dynamically injected into the LLM system context at runtime

Business value

• Enforces firm-wide house style and attorney-specific drafting habits
• Reduces “micro-friction” edits and repeated preference corrections
• Supports consistent examiner strategy posture across matters

4) One-Click EPO Prior Art Ingestion

PatentFlow integrates EPO retrieval to support:

• automated ingestion of cited prior art (e.g., D1/D2 full text)
• reduced manual copy/paste and document hunting
• faster turnaround from Office Action to structured analysis

Business value

• Cuts administrative time
• Increases completeness and consistency of cited-document context
• Improves auditability of the evidence basis used in analysis

---

Quick Start

Option A — Docker (recommended for reproducibility)

1. Configure environment:

• Copy .env.example → .env
• Set NEXT_PUBLIC_API_BASE_URL, REDIS_URL, and LLM configuration as needed

2. Start services:

docker compose up --build

Typical services:

• frontend (Next.js UI)
• api (FastAPI gateway)
• redis (broker + result backend)
• worker (Celery worker, local LLM calls)

Option B — Manual (local development)

1) Backend (FastAPI)

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

export REDIS_URL=redis://localhost:6379/0
uvicorn src.api:app --host 0.0.0.0 --port 8000

2) Redis

redis-server

3) Celery Worker

export REDIS_URL=redis://localhost:6379/0
celery -A src.celery_app.celery_app worker -l info --concurrency=1 --prefetch-multiplier=1

4) Frontend (Next.js)

cd frontend
npm install
npm run dev

Open:

• UI: http://localhost:3000
• API: http://localhost:8000/health

---

API Overview (Selected)

• POST /api/generate
  • Runs the async pipeline (Celery) for claim chart + verification + draft outputs
• GET /api/status/{task_id}
  • Poll for progress and results
• GET /api/memory/{attorney_id}
  • Retrieve stored preference string
• POST /api/memory/add
  • Append a new preference rule for an attorney profile
• POST /api/generate-chart
  • Deterministic + LLM-assisted chart generation with optional attorney_id

---

Security & Privacy Posture

• Designed for offline and air-gapped operation
• Local persistence only (SQLite)
• No dependency on third-party analytics, telemetry, or cloud inference for core workflows

---

Roadmap (Prioritized for Firm Integration)

• RAG with ChromaDB (Depth)
  Local retrieval over firm-approved corpora (e.g., standards, prior OA templates) to improve long-document reasoning while controlling hallucination risk.

• .docx Export Workflow (Adoption)
  Export claim charts and drafted responses into Word with firm formatting and review conventions.

• SSE Streaming (UX)
  Upgrade from polling to server-sent events for long-running generation, keeping the interface calm and traceable.

• Policy Packs (Governance)
  Versioned preference bundles per firm/practice group to standardize style and examiner strategy guidance.

---

License / Intended Use

This repository is intended for professional evaluation and internal deployment patterns.

For production firm deployments, additional hardening (audit logs, access controls, document storage policies) is recommended.