Visual Routing Editor

The Eluvix Visual Routing Editor (accessible at /editor) is an interactive, graph-based canvas powered by React Flow for designing, inspecting, and compiling multimodal routing topologies, agent gateways, and security guardrails.

Canvas Interface & Usability

The editor provides a modern, dark-slate visual development environment with ergonomic design enhancements:

Ergonomic Node Handles: Enlarged, color-coded grab handles with extended hitboxes for effortless connection wiring.
High-Visibility SVG Wires: Thicker bezier connection wires with distinct color-coding based on the source port's operational layer.
Dynamic Port Handles: Nodes such as the Task Router and Switchyard dynamically project multiple semantic output ports (e.g. #intent-code, #intent-reasoning, #tier-local, #tier-cloud).
Interactive In-Card Controls: Direct on-canvas toggles, sliders, and inputs for immediate adjustments without opening deep modal menus.
Properties Inspector Panel: Selecting any node on the canvas opens a dedicated inspector panel on the right sidebar for fine-grained configuration.

[ Client Entrypoint ]
         │
         ▼
[ Privacy Guard Node ] ──► In-flight PII Masking (Token / Asterisk / Hash)
         │
         ▼
[ Semantic Guard Node ] ──► Prompt Injection & Jailbreak Classifier
         │
         ▼
[ Task Router Node ] ──► Semantic Intent Dispatcher
         │
         ├───► #intent-code ──────► [ Coding Agent Node ]
         ├───► #intent-reasoning ─► [ Google Antigravity AGY ]
         └───► #intent-default ───► [ Switchyard Router Node ]
                                            │
                                 ┌──────────┴──────────┐
                                 ▼                     ▼
                       [ Model Gateway ]        [ Cloud Providers ]
                         (Local vLLM /           (Billama / OpenAI /
                          Ollama Cache)           Claude Direct)

Comprehensive Node Catalog

The Routing Editor organizes nodes into functional categories:

1. Security Nodes

Privacy Guard (privacyGuardNode):
- Enforces in-flight PII redaction using token ([REDACTED_API_KEY]), asterisk (sk-••••), or cryptographic hash ([HASH:e3b0...] via SHA-256).
- Built-in entity filters: API keys, SSNs, credit cards, emails, phone numbers, IP addresses.
- Exempted Whitelist: Domain and token exemptions (e.g. internal.corp, localhost).
- Custom Regex: Configurable user patterns with named identifiers.
- Audit Logging: Optional logging of redacted fragments for security auditing.
Semantic Guard (semanticGuardNode):
- Embedding classifier firewall models (DeBERTa-v3, BGE-Small, MiniLM).
- Sensitivity slider (0.50–0.99).
- Enforcement actions: block (HTTP 400), strip (sanitized pass-through), divert (quarantine sink), warn (audit only).
- Active threat vectors: Instruction overrides, DAN jailbreaks, system prompt leaks, obfuscated attacks.
- Custom banned phrases blacklist.

2. Routing & Intent Dispatchers

Switchyard Router (switchyardNode):
- Layer 7 adaptive balancer supporting 5 routing strategies (cost-complexity, cascade-cost, lowest-cost, lowest-latency, round-robin).
- Configurable latency spillover circuit breaker (spilloverThresholdMs).
- Upstream network timeout (timeoutMs) and retry counts.
- Token complexity boundaries and automatic code escalation.
- Financial spending cap (maxCostPerRequest).
Task Router (taskRouterNode):
- Dispatches queries based on semantic intent using Heuristic Regex, FastEmbed Similarity, or Micro-LLM classification.
- Dynamic intent target handles: #intent-code, #intent-reasoning, #intent-fast, #intent-tools, #intent-default.
- Routing modes: Balanced, Lowest Cost, Max Speed, Capability Match.

3. Gateways & Proxies

Model Gateway (modelGatewayNode):
- OpenAI-compatible Bifrost adapter for local and upstream models.
- In-memory semantic response cache with configurable TTL (seconds) and similarity threshold (X-Eluvix-Cache: HIT).
- Sliding-window RPM rate limiter with automatic HTTP 429 enforcement.
- Fallback strategies: Cheaper Tier, Same-Tier Alternative, Fail Fast.
- Global Enterprise System Prompt Prepend.
Agent Gateway (agentGatewayNode):
- Secures Model Context Protocol (MCP) and Agent-to-Agent (A2A) tool invocations.
- Access policies: restrictive (whitelist only), permissive (internal VPC), open (public registry).
- Tool filter modes: all, whitelist, blacklist.
- Output PII sanitization: Scrubs sensitive data from tool returns before feeding to LLMs.
- OpenBao Secret Mapping: Directly maps encrypted API keys into tool runtime environments.

4. Tools & Observability

MCP Tool (mcpToolNode):
- Connects external MCP servers via SSE or stdio endpoints.
- Execution modes: auto (immediate), confirm (human-in-the-loop), dry-run (simulation).
- Configurable execution timeouts.
Observability Node (observabilityNode):
- Multi-exporter telemetry engine: Embedded SQLite, OpenTelemetry Collector (OTLP/gRPC), Langfuse LLM Tracing, Prometheus.
- Trace sampling rate slider (1–100%).
- PII scrubbing from recorded traces.
- Full request input/output capture toggle.
- SLA latency alert threshold (ms).

5. Agents & Runtimes

Custom Agent (agentNode):
- Autonomous persona definitions with system instructions and temperature controls.
- Primary capability specialization: General, Coding, Research, Orchestration.
- Persistent multi-turn memory toggle.
Local Engine Nodes (lemonadeNode, vllmNode, ollamaNode):
- Local GPU/NPU-accelerated runtimes.
- Priority scoring, max parallel concurrency, and context window limits.
Cloud Providers & Agent CLI Nodes:
- Billama Cloud, OpenAI Direct, Anthropic Claude Direct, Google Antigravity AGY, OpenAI Codex, Claude Code.

Multi-Target Compiler (`compiler.ts`)

Clicking "Compile & Save" converts visual graph connections into 10 synchronized configuration artifacts:

Compiled configurations are saved to SQLite via POST /api/v1/config and immediately hot-reloaded without service downtime.