How MEND-X Thinks.
Live Architecture Flowchart & Execution SimulatorInteractive v1.2
Integrated System Topology & Live Packet Trace
Select a live scenario or enter a custom query to watch data packets traverse the actual FastAPI, pgvector, Cross-Encoder, and 3-Tier AI router pipeline in real time.
Custom Query Probe:
Active Query:Spindle overload alarm E-402 on Haas VF-2 CNC
Stage:Ready to Execute
SLA Target:1,240 ms
Evidence:0.89
Route:GPT-OSS 20B (Groq Fast)
Step-by-Step Architecture PipelineClick any node to inspect source code & payload
Tier 1 · Client & Ingestion< 8ms
Tier 2 · Query Classification & 1536-dim Embedding~30ms
Tier 3 · Hybrid Retrieval (pgvector HNSW + BM25) & Reranking40–80ms
Tier 4 · Zero-Hallucination Guardrails & Cutoff Filter< 5ms
Tier 5 · Adaptive Model Cascade RouterMini <100ms | 20B 1–2s | 120B 2–4s
Tier 6 · Citation Hydration & Sub-second SSE Delivery< 10ms
Zero-Hallucination Guardrail
Evidence Validator & Refusal Circuit
Calculates total evidence sufficiency score. If score is below 0.72 threshold, triggers refusal circuit immediately.
Engine:Evidence Cutoff (Threshold ≥ 0.72)
Source:backend/app/services/rag/evidence_validator.py
class EvidenceValidator:
def validate(self, chunks: list[Chunk]) -> ValidationResult:
if not chunks:
return ValidationResult(is_sufficient=False, evidence_score=0.0)
score = calculate_composite_evidence(chunks)
if score < self.threshold:
return ValidationResult(is_sufficient=False, evidence_score=score)
return ValidationResult(is_sufficient=True, evidence_score=score)Scenario Expected Resultsuccess
Direct error code lookup. High retrieval confidence (0.89), passes disambiguation, routes to FORGE for verified 3-step repair.
| Phase | Latency | Details |
|---|---|---|
| Query Ingestion | <10ms | API gateway processing |
| pgvector ANN Search | 40–80ms | Cosine NN search (1M vectors) |
| Nord Inference | <100ms | Nord (Groq LPU) |
| Forge Inference | 1–2s | Forge (Fast Diagnostics) |
| Apex Inference | 2–4s | Apex (Deep Reasoning) |
| Response Serialization | <5ms | JSON + streaming overhead |
Target SLA: Mini <100ms, 20B 1–2s, 120B 2–4s. Caching and model selection ensure sub-second median for 92% of queries.
| Scenario | Avg Latency | P99 | Throughput | Accuracy |
|---|---|---|---|---|
| Error Code Lookup (Nord) | 67ms | 142ms | 14,900 q/s | 99.2% |
| Multi-Step Procedure (Forge) | 1.24s | 2.1s | 240 tok/s | 98.4% |
| Root Cause Analysis (Apex) | 2.6s | 4.2s | 180 tok/s | 99.4% |
| Cold Start (Cache Miss) | 512ms + model latency | 1.2s + model latency | N/A | N/A |
Precision Built
Into the Core.
Zero-hallucination RAG, tenant isolation, multi-tier routing, and compliance-ready deployment. Engineering that matches industrial demands.