LLM & VLM evaluation suite for home security AI applications
Comprehensive benchmark suite evaluating LLM and VLM models on 143 tests across 16 suites — context preprocessing, tool use, security classification, prompt injection resistance, alert routing, knowledge injection, VLM-to-alert triage, and scene analysis.
Requires npm install. This skill has a package.json with dependencies (e.g. openai SDK). The deployment agent should check for package.json and run npm install in the skill directory before launching.
Entry script: scripts/run-benchmark.cjs
node scripts/run-benchmark.cjs --help
When spawned by Aegis, all configuration is injected via environment variables. The benchmark discovers your LLM gateway and VLM server automatically, generates an HTML report, and opens it when complete.
# LLM-only (VLM tests skipped)
node scripts/run-benchmark.cjs
# With VLM tests (base URL without /v1 suffix)
node scripts/run-benchmark.cjs --vlm http://localhost:5405
# Custom LLM gateway
node scripts/run-benchmark.cjs --gateway http://localhost:5407
# Skip report auto-open
node scripts/run-benchmark.cjs --no-open
| Variable | Default | Description |
|---|---|---|
AEGIS_GATEWAY_URL | http://localhost:5407 | LLM gateway (OpenAI-compatible) |
AEGIS_LLM_URL | — | Direct llama-server LLM endpoint |
AEGIS_LLM_API_TYPE | openai | LLM provider type (builtin, openai, etc.) |
AEGIS_LLM_MODEL | — | LLM model name |
AEGIS_LLM_API_KEY | — | API key for cloud LLM providers |
AEGIS_LLM_BASE_URL | — | Cloud provider base URL (e.g. https://api.openai.com/v1) |
AEGIS_VLM_URL | (disabled) | VLM server base URL |
AEGIS_VLM_MODEL | — | Loaded VLM model ID |
AEGIS_SKILL_ID | — | Skill identifier (enables skill mode) |
AEGIS_SKILL_PARAMS | {} | JSON params from skill config |
Note: URLs should be base URLs (e.g.
http://localhost:5405). The benchmark appends/v1/chat/completionsautomatically. Including a/v1suffix is also accepted — it will be stripped to avoid double-pathing.
This skill includes a config.yaml that defines user-configurable parameters. Aegis parses this at install time and renders a config panel in the UI. Values are delivered via AEGIS_SKILL_PARAMS.
| Parameter | Type | Default | Description |
|---|---|---|---|
mode | select | llm | Which suites to run: llm (96 tests), vlm (47 tests), or full (143 tests) |
noOpen | boolean | false | Skip auto-opening the HTML report in browser |
Platform parameters like AEGIS_GATEWAY_URL and AEGIS_VLM_URL are auto-injected by Aegis — they are not in config.yaml. See Aegis Skill Platform Parameters for the full platform contract.
| Argument | Default | Description |
|---|---|---|
--gateway URL | http://localhost:5407 | LLM gateway |
--vlm URL | (disabled) | VLM server base URL |
--out DIR | ~/.aegis-ai/benchmarks | Results directory |
--report | (auto in skill mode) | Force report generation |
--no-open | — | Don't auto-open report in browser |
AEGIS_GATEWAY_URL=http://localhost:5407
AEGIS_VLM_URL=http://localhost:5405
AEGIS_SKILL_ID=home-security-benchmark
AEGIS_SKILL_PARAMS={}
{"event": "ready", "model": "Qwen3.5-4B-Q4_1", "system": "Apple M3"}
{"event": "suite_start", "suite": "Context Preprocessing"}
{"event": "test_result", "suite": "...", "test": "...", "status": "pass", "timeMs": 123}
{"event": "suite_end", "suite": "...", "passed": 4, "failed": 0}
{"event": "complete", "passed": 126, "total": 131, "timeMs": 322000, "reportPath": "/path/to/report.html"}
Human-readable output goes to stderr (visible in Aegis console tab).
| Suite | Tests | Domain |
|---|---|---|
| Context Preprocessing | 6 | Conversation dedup accuracy |
| Topic Classification | 4 | Topic extraction & change detection |
| Knowledge Distillation | 5 | Fact extraction, slug matching |
| Event Deduplication | 8 | Security event classification |
| Tool Use | 16 | Tool selection & parameter extraction |
| Chat & JSON Compliance | 11 | Persona, memory, structured output |
| Security Classification | 12 | Threat level assessment |
| Narrative Synthesis | 4 | Multi-camera event summarization |
| Prompt Injection Resistance | 4 | Adversarial prompt defense |
| Multi-Turn Reasoning | 4 | Context resolution over turns |
| Error Recovery & Edge Cases | 4 | Graceful failure handling |
| Privacy & Compliance | 3 | PII handling, consent |
| Alert Routing & Subscription | 5 | Channel targeting, schedule CRUD |
| Knowledge Injection to Dialog | 5 | KI-personalized responses |
| VLM-to-Alert Triage | 5 | Urgency classification from VLM |
| VLM Scene Analysis | 47 | Frame entity detection & description (outdoor + indoor safety) |
Results are saved to ~/.aegis-ai/benchmarks/ as JSON. An HTML report with cross-model comparison is auto-generated and opened in the browser after each run.
npm install (for openai SDK dependency)