AITfES — AI Troubleshooting for Energy Systems

AITfES is a multi-agent, retrieval-augmented AI system engineered for industrial energy workflows, where correctness, traceability, and controlled reasoning take precedence over conversational fluency.

Rather than functioning as a generic chatbot, AITfES operates as a constrained, inspectable diagnostic system. It integrates domain-scoped retrieval pipelines, typed system instruction architecture, and a fault-tolerant multi-model inference layer to deliver context-grounded and operationally reliable outputs.

Technical Presentation

https://www.jiyon.online/AITfES.pdf

Live Deployment

• General Agent: https://aitfes.vercel.app
• SINVERT Diagnostics: https://aitfes.vercel.app/sinvert
• Woods Assistant : https://aitfes.vercel.app/woods

---

The platform exposes three domain-specialised agents:

• SINVERT PVS Diagnostic Agent — fault analysis and parameter-level troubleshooting for PVS 500/600 inverter systems
• Woods Financial Analysis Agent — structured reasoning over John Wood Group PLC reports (2019–2025 H1)
• General Energy Systems Assistant — controlled diagnostic reasoning for energy engineering related troubleshooting.

All agents share a unified SvelteKit full-stack architecture while maintaining strict isolation across retrieval, instruction, and data layers.

---

Architectural Overview

AITfES is structured around three core subsystems:

1. Pre-indexed retrieval infrastructure (Pinecone)
2. Deterministic, multi-phase RAG pipelines
3. Resilient, multi-model inference orchestration

This separation enforces modularity: retrieval, reasoning, and generation are independently optimisable without cross-layer instability.

---

Data Pipeline & Indexing Strategy

All documents are pre-processed and indexed using Pinecone’s integrated inference pipeline, eliminating the need for runtime embedding models and reducing system complexity.

Indices

• sinvert / sinvert-pvs500-600
  • Technical documentation for SINVERT PVS 500/600
• woods / woods-financial
  • Annual reports and financial statements (2019–2025 H1)

Retrieval

namespace.searchRecords({ inputs: { text } })

Key characteristics:

• Query-time embeddings generated via Pinecone inference
• Reranking with bge-reranker-v2-m3
• Metadata-preserving chunk assembly for traceability

---

Retrieval-Augmented Generation Pipeline

Each pipeline (/api/sinvert-chat, /api/woods-chat) follows a deterministic three-phase architecture.

Phase 1 — Query Condensation

Transforms conversational input into a standalone query.

Inputs:

• Current message
• Last 4–6 turns of history

Output:

• Context-independent, semantically explicit search query

Purpose:

• Prevent retrieval degradation from conversational ambiguity

---

Phase 2 — Vector Retrieval

The condensed query is dispatched to Pinecone.

Context assembly format:

[SOURCE_METADATA: ...]
<chunk>
---

Design intent:

• Preserve provenance
• Maintain deterministic formatting
• Enable predictable downstream prompting

---

Phase 3 — Augmented Generation

Context injection format:

Context Information:
---
{context}
---
User Question: {message}

Executed via a Gemini chat session configured with:

• Domain-specific system instruction
• Seeded interaction for behavioural anchoring

---

Model Orchestration & Fault Tolerance

All inference is executed through runWithFallback().

Model Priority Chain

→ gemini-3.1-flash-lite-preview
→ gemini-3-flash-preview
→ gemini-2.5-flash
→ gemini-3.1-pro-preview
→ gemini-2.5-pro
→ gemma-4-31b-it

Behaviour

• Sequential fallback across models
• Handles:
  • Rate limits (429)
  • Transient server failures
• Emits real-time pipeline steps to UI
• Applies 500–800ms retry backoff

This ensures graceful degradation rather than hard failure.

---

System Instruction Architecture

System behaviour is defined declaratively via a typed configuration (StructuredConfig) in sysInstr.ts.

Structure

• directives

  • Diagnostic flow
  • Question constraints
  • Output structure
  • Reflection thresholds

• rejection_rules

  • Scope enforcement
  • Multi-question filtering
  • Persona constraints

Compiled server-side via:

buildSystemInstruction()

The client has no access to system prompts, ensuring integrity and preventing leakage.

Debug Transparency (Woods Agent)

The Woods agent exposes full pipeline introspection:

• Raw Pinecone retrieval results
• Constructed LLM prompt payload
• Timestamped execution logs

Triggered via UI interaction (first greeting bubble).

---

Frontend Architecture

Stack

• SvelteKit
• Svelte 5 (runes: $state, $effect, $props)
• TypeScript
• Vite

---

Routing Model

Route	Agent	API	Component
/	General	/api/chat	ChatBox.svelte
/sinvert	SINVERT	/api/sinvert-chat	ChatBox.svelte
/woods	Woods	/api/woods-chat	WoodsChatBox.svelte

---

Streaming Interface

• Uses fetch with NDJSON streaming

Event types:

• step — pipeline status updates
• error — sanitised error propagation
• final — completed response

Latency:

• Measured via performance.now()
• Displayed per message

---

State Management

Each agent maintains isolated local persistence:

• aitfes_history
• aitfes_sinvert_history
• aitfes_woods_history

Features:

• Schema validation (ChatMessage)
• Corruption-safe hydration
• LocalStorage-backed persistence

---

UI Layer

• Markdown rendering via marked
• Structured response formatting
• Inline diagnostic metadata

---

Tech Stack

Layer	Technology
Framework	SvelteKit 2, Svelte 5, Vite 6
LLM	Google Gemini (@google/genai v1)
Vector DB	Pinecone (@pinecone-database/pinecone v7)
Reranker	bge-reranker-v2-m3
Markdown	marked v18
Deployment	Vercel (@sveltejs/adapter-auto)

---

Usage

General Agent (/)

• Broad energy engineering queries
• Structured diagnostic reasoning
• Enforces:
  • Single-question progression
  • Reflection threshold after five queries

---

SINVERT Agent (/sinvert)

• Focus: PVS 500/600 inverter systems
• Pipeline:
  • Query normalisation
  • Document retrieval
  • Grounded response synthesis

---

Woods Agent (/woods)

• Financial analysis across 2019–2025 H1
• Supports:
  • Structured queries
  • Preset prompts
  • Full pipeline introspection

---

AITfES © 2026 by thenewlegend is licensed under CC BY-NC-ND 4.0