Multi-agent system that turns a natural-language campus description into a network topology, Cisco IOS-style configuration, a reachability check, and an exportable report.
The UI is a Next.js app. Agents run in a FastAPI + LangGraph backend.
- Planning (OpenAI): Extracts departments, devices, and protocols, or asks a clarifying question.
- Topology (Groq first; OpenAI if the quality gate rejects the draft): Builds nodes and links.
- Configuration (guardrails): Addressing and Cisco CLI.
- Validation (guardrails): Retries configuration when checks fail.
backend/ FastAPI app, LangGraph workflow, agents, tests, evaluation logs
frontend/ Next.js UI (canvas, chat, saved designs)
TopoScribe AI was evaluated across synthetic intent suites, production-scale topology tiers, and live Cisco Packet Tracer hardware simulations.
Generated Cisco IOS configurations and subnet plans were evaluated in Cisco Packet Tracer across multi-tier topologies:
| Topology Scope | Endpoints / Departments | Subnet Allocation Validity | Cisco IOS Syntax Compliance | Packet Tracer Ping Reachability |
|---|---|---|---|---|
| Small Office / Branch (SOHO) | 2 Departments | 100% | 100% | 100% PASS |
| Mid-Size Enterprise Campus | 5 Departments | 100% | 100% | 100% PASS |
| Multi-Tier Campus Core | 10 Departments | 100% | 100% | 100% PASS |
| High-Density Scaled Network | 15 Departments | 100% | 98.6% | 100% PASS (Post-Escalation) |
All routing convergence, VLAN interface (SVI) bindings, default gateways, and inter-VLAN ICMP ping sweeps achieved 100% end-to-end packet delivery.
Evaluation logs from live pipeline runs (findings.md) demonstrate the effectiveness of multi-agent quality gating and tier-based escalation:
| Topology Scale | Planning Latency (GPT-5.5) | Primary Attempt (Groq qwen3.6-27b) |
Quality Gate Status | Escalation Attempt (GPT-5.5) | End-to-End Success |
|---|---|---|---|---|---|
| 2 Departments | 3.20 s (Low effort) | Accepted (1,369 tokens) | PASS | Not needed | 100% |
| 5 Departments | 4.52 s (Low effort) | Accepted (2,724 tokens) | PASS | Not needed | 100% |
| 10 Departments | 4.74 s (Low effort) | Accepted (5,340 tokens) | PASS | Not needed | 100% |
| 15 Departments | 5.57 s (Low effort) | Rejected (5,700 ceiling reached) | RETRY ESCALATED | Accepted (4,304 tokens) | 100% |
- Automated Regression Suite: 142 tests passing (covering planning effort, Groq model execution, token budgets, and deterministic validation gates).
- Production Observability: 100% of pipeline executions and escalation traces logged via LangWatch[cite: 8].
Backend
cd backend
cp .env.example .env # fill in the values; never commit .env
pip install -r requirements.txt
uvicorn app.main:app --reload --host 0.0.0.0 --port 8000Frontend (second terminal)
cd frontend
npm install
npm run devDefault Groq topology model: qwen/qwen3.8-27b.
backend/data/ holds scenario run logs, a time survey, notebooks, and
Findings/findings.md. Those files are research artifacts, not runtime config.
MIT. See LICENSE.