🧠 KnowledgeBrain API

AI for Industrial Knowledge Intelligence — Unified Asset & Operations Brain

One unified intelligence layer across all your industrial documents. Query maintenance records, engineering drawings, compliance procedures, and incident reports in natural language — in seconds, not hours.

🚀 Quick Start · 📐 Architecture · 🔌 API Reference · 📊 Evaluation · 👥 Team

📌 The Problem

Industrial facilities in India and globally are drowning in documents — yet paradoxically starved of actionable knowledge at the moment decisions need to be made.

Statistic	Source
35% of industrial professionals' working hours lost to information search	McKinsey 2024
7–12 disconnected document systems in the average large Indian plant	NASSCOM-EY
18–22% of unplanned downtime events attributable to knowledge fragmentation	BIS Research
25% of experienced engineers retiring within 10 years, taking tacit knowledge with them	Industry estimate

A maintenance technician diagnosing a pump failure cannot instantly cross-reference the OEM manual, the last three inspection reports, the work order history, and the relevant OISD procedure in a single query. Each document lives in a different system, in a different format, accessible only to a different department. KnowledgeBrain solves this.

✨ Features

1. 📄 Universal Document Ingestion & Knowledge Graph

Ingest every document type found in an industrial facility into a single queryable intelligence layer:

PDFs — engineering specs, MSDS, regulatory submissions, operating manuals
P&IDs — plant and instrumentation diagrams parsed via computer vision (YOLOv8)
Scanned forms — inspection records, permit-to-work logs, maintenance sign-offs (OCR via Azure Document Intelligence)
Spreadsheets — maintenance history, calibration records, spare parts inventories
Email archives — informal operational knowledge, incident alerts, project correspondence

Entities (equipment tags, process parameters, regulatory references, personnel, dates, failure codes) are extracted and connected in a Neo4j property knowledge graph that auto-updates as new documents arrive.

2. 🤖 Expert Knowledge Copilot

A RAG-powered conversational AI that answers operational, maintenance, and engineering queries against the full document corpus:

Source-cited — every answer links to originating documents with page references
Confidence-scored — communicates uncertainty, never false certainty
Multi-language — supports English, Hindi (हिन्दी), Telugu (తెలుగు), Tamil (தமிழ்)
Offline-capable — core query works with local model cache in low-connectivity zones
Mobile-first — field technicians query in the field, not engineers at desktops

Example queries handled:

"What is the last recorded valve torque for V-204B and when was it calibrated?"
"Show me the lockout/tagout procedure for the boiler feed pump at Train 2."
"What were the findings from the last OISD audit and which NCRs are still open?"
"What caused the similar vibration issue on P-101A in 2022 and what was the fix?"

3. 🔧 Maintenance Intelligence & RCA Agent

An agentic system (LangGraph) that fuses work order history, equipment failure records, OEM manuals, and real-time readings to:

Generate predictive maintenance recommendations based on failure pattern analysis
Provide structured Root Cause Analysis (RCA) using the 5-Whys methodology
Optimise maintenance schedules against operational windows
Surface cross-equipment failure patterns invisible to individual maintenance teams
Reduce MTTR by pre-staging the right information before the technician arrives

4. 🛡️ Quality & Regulatory Compliance Intelligence

An agentic compliance layer mapping regulatory requirements (Factory Act, OISD, PESO, BIS) against current procedures and equipment states:

Continuously identifies compliance gaps before audits find them
Auto-generates compliance evidence packages for Factory Inspector or OISD audits
Flags quality deviations in real time and routes to the responsible owner
Live regulatory change tracker — when OISD publishes an update, affected procedures are flagged immediately

5. 📚 Lessons Learned & Failure Intelligence Engine

Institutional memory across the organisation's entire history of incidents and near-misses:

Identifies systemic patterns no single investigation would detect
Proactively pushes warnings when current conditions match pre-failure patterns
Preserves tacit knowledge of retiring engineers by structuring their incident narratives into the knowledge graph
Delivers a "similar incidents" briefing the moment a new incident is logged

📐 Architecture

┌─────────────────────────────────────────────────────────────┐
│            LAYER 5 — Multi-Channel Interface                │
│         Web (React)  ·  Mobile (React Native)  ·  API       │
├─────────────────────────────────────────────────────────────┤
│               LAYER 4 — Application Layer                   │
│   Copilot  ·  Maintenance Agent  ·  Compliance Agent        │
│                   ·  Lessons Learned Engine                 │
├─────────────────────────────────────────────────────────────┤
│            LAYER 3 — RAG + Agentic Reasoning                │
│       LlamaIndex  ·  Qdrant  ·  LangGraph Agents            │
├─────────────────────────────────────────────────────────────┤
│          LAYER 2 — Knowledge Graph & Entity Store           │
│      Neo4j Property Graph  ·  Equipment-Procedure-Risk      │
├─────────────────────────────────────────────────────────────┤
│         LAYER 1 — Document Ingestion & Processing           │
│   PDF  ·  P&ID (YOLOv8)  ·  OCR  ·  Spreadsheet  ·  Email  │
└─────────────────────────────────────────────────────────────┘

Data Flow

Raw Documents ──► Ingestion Pipeline ──► Chunking & Entity Extraction
                                                │
                              ┌─────────────────┴─────────────────┐
                              ▼                                   ▼
                        Qdrant Vector DB                   Neo4j Graph DB
                        (semantic search)              (entity relationships)
                              │                                   │
                              └─────────────────┬─────────────────┘
                                                │
                                    Hybrid Retrieval Engine
                                                │
                                    Claude claude-sonnet-4-6 (LLM)
                                                │
                                     Structured Response
                              (answer + sources + confidence + citations)

🛠️ Technology Stack

Layer	Technology	Purpose
LLM Backend	Claude `claude-sonnet-4-6` (Anthropic)	Primary reasoning, RAG synthesis, multi-language
RAG Framework	LlamaIndex	Document ingestion pipelines, chunking, retrieval
Vector Database	Qdrant	High-performance semantic search
Knowledge Graph	Neo4j	Equipment-document-risk entity relationships
OCR & Doc Intel	Azure Document Intelligence	Scanned PDFs, forms, P&IDs
Computer Vision	YOLOv8 (custom fine-tune)	P&ID symbol detection and tag extraction
Agentic Framework	LangGraph	Multi-agent orchestration
Backend API	FastAPI (Python 3.11)	REST + WebSocket API layer
Frontend	React + TailwindCSS	Engineer and manager dashboards
Mobile	React Native (Expo)	Field technician mobile interface
Infrastructure	Docker + Kubernetes	Container-native, cloud-agnostic deployment
Storage	MinIO + PostgreSQL + Redis	Documents, metadata, caching
Auth	Keycloak	Role-based access control

📁 Project Structure

KnowledgeBrainAPI/
├── app/
│   ├── main.py                     # FastAPI entry point
│   ├── config.py                   # Settings via pydantic-settings
│   ├── routers/
│   │   ├── ingest.py               # POST /ingest
│   │   ├── query.py                # POST /query
│   │   ├── copilot.py              # POST /copilot/chat
│   │   ├── maintenance.py          # POST /maintenance/predict, /rca
│   │   ├── compliance.py           # POST /compliance/gaps, /audit-package
│   │   ├── lessons.py              # POST /lessons/analyze-incident
│   │   └── health.py               # GET  /health
│   ├── services/
│   │   ├── ingestion/
│   │   │   ├── pdf_processor.py
│   │   │   ├── pid_processor.py    # YOLOv8 P&ID parser
│   │   │   ├── spreadsheet_processor.py
│   │   │   └── email_processor.py
│   │   ├── rag/
│   │   │   ├── embedder.py
│   │   │   ├── retriever.py        # Hybrid semantic + graph retrieval
│   │   │   └── synthesizer.py      # Claude-powered answer generation
│   │   ├── graph/
│   │   │   ├── neo4j_client.py
│   │   │   ├── entity_extractor.py
│   │   │   └── graph_builder.py
│   │   └── agents/
│   │       ├── maintenance_agent.py   # LangGraph agent
│   │       ├── compliance_agent.py    # LangGraph agent
│   │       └── lessons_agent.py       # LangGraph agent
│   └── models/
│       ├── document.py
│       ├── query.py
│       └── alert.py
├── frontend/                       # React + TailwindCSS web app
├── mobile/                         # React Native (Expo) field app
├── eval/
│   ├── benchmark_questions.json    # 20 domain-expert questions
│   ├── ground_truth.json           # Entity extraction ground truth
│   ├── test_entity_extraction.py
│   ├── test_query_quality.py
│   ├── test_graph_linkage.py
│   ├── test_performance.py
│   └── run_eval.py                 # Master evaluation runner
├── data/
│   └── synthetic/                  # Demo dataset (OISD reports, sample P&IDs)
├── scripts/
│   └── generate_demo_data.py
├── docker-compose.yml
├── Dockerfile
├── requirements.txt
├── .env.example
└── README.md

🚀 Quick Start

Prerequisites

Python 3.11+
Node.js 18+
Docker & Docker Compose
An Anthropic API key
Azure Document Intelligence resource (for OCR)

1. Clone the repository

git clone https://github.com/SakshamDevloper/KnowledgeBrainAPI.git
cd KnowledgeBrainAPI

2. Set up environment variables

cp .env.example .env
# Edit .env and fill in your API keys (see Environment Variables below)

3. Run with Docker Compose (recommended)

docker-compose up --build

This starts: FastAPI backend · Qdrant · Neo4j · Redis · MinIO

App will be available at http://localhost:8000 API docs at http://localhost:8000/docs

4. Run for development (without Docker)

# Backend
python -m venv venv
source venv/bin/activate          # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

# Frontend (separate terminal)
cd frontend
npm install
npm run dev                        # http://localhost:3000

🔑 Environment Variables

Create a .env file in the project root:

# ── Anthropic ──────────────────────────────────
ANTHROPIC_API_KEY=sk-ant-...

# ── Qdrant ─────────────────────────────────────
QDRANT_HOST=localhost
QDRANT_PORT=6333
QDRANT_COLLECTION_NAME=knowledgebrain_docs

# ── Neo4j ──────────────────────────────────────
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your_password_here

# ── Azure Document Intelligence ────────────────
AZURE_FORM_RECOGNIZER_ENDPOINT=https://your-resource.cognitiveservices.azure.com/
AZURE_FORM_RECOGNIZER_KEY=your_azure_key

# ── Redis ──────────────────────────────────────
REDIS_HOST=localhost
REDIS_PORT=6379

# ── MinIO (Object Storage) ─────────────────────
MINIO_ENDPOINT=localhost:9000
MINIO_ACCESS_KEY=minioadmin
MINIO_SECRET_KEY=minioadmin
MINIO_BUCKET=knowledgebrain

# ── App Settings ───────────────────────────────
APP_ENV=development
LOG_LEVEL=INFO
MAX_UPLOAD_SIZE_MB=100
DEFAULT_TOP_K=10
CONVERSATION_TTL_HOURS=24

🔌 API Reference

Document Ingestion

Method	Endpoint	Description
`POST`	`/ingest/pdf`	Ingest a PDF document
`POST`	`/ingest/spreadsheet`	Ingest Excel/CSV maintenance data
`POST`	`/ingest/email`	Ingest email archive (.eml / .txt)
`POST`	`/ingest/batch`	Batch ingest a folder of documents
`GET`	`/ingest/status/{job_id}`	Check ingestion job status

Copilot Chat

Method	Endpoint	Description
`POST`	`/copilot/chat`	Send a query, receive answer + sources
`GET`	`/copilot/history/{conversation_id}`	Retrieve conversation history
`DELETE`	`/copilot/history/{conversation_id}`	Clear conversation

Example request:

POST /copilot/chat
{
  "message": "What caused the vibration failure on P-101A in 2022?",
  "language": "en",
  "equipment_context": ["P-101A"],
  "conversation_id": "optional-session-id"
}

Example response:

{
  "answer": "The 2022 vibration failure on P-101A was caused by...",
  "sources": [
    {"filename": "RCA_P101A_2022.pdf", "page": 4, "relevance": 0.94}
  ],
  "confidence": "HIGH",
  "follow_up_suggestions": ["What was the resolution?", "Is this failure mode common?"],
  "safety_flag": false,
  "conversation_id": "conv_abc123"
}

Maintenance Intelligence

Method	Endpoint	Description
`POST`	`/maintenance/predict`	Predictive maintenance for an equipment tag
`POST`	`/maintenance/rca`	Run root cause analysis for an incident
`GET`	`/maintenance/schedule/{equipment_tag}`	Get maintenance schedule

Compliance

Method	Endpoint	Description
`POST`	`/compliance/gaps`	Detect compliance gaps (all or by equipment)
`POST`	`/compliance/audit-package`	Generate audit evidence package
`POST`	`/compliance/check-regulation-change`	Check impact of a regulation update

Lessons Learned

Method	Endpoint	Description
`POST`	`/lessons/analyze-incident`	Analyze incident, find similar past events
`POST`	`/lessons/capture-expert-knowledge`	Capture retiring engineer's knowledge
`GET`	`/lessons/alerts`	Get active proactive alerts

📊 Evaluation Metrics

Run the full evaluation suite:

cd eval
python run_eval.py
# Outputs: evaluation_summary.json

Metric	Target	How Measured
Entity Extraction Accuracy	> 90% F1	Equipment tags, process params, regulatory refs across 5 doc types
Query Answer Quality	> 85% relevance	Claude-as-judge on 20 domain-expert benchmark questions
Knowledge Graph Linkage	> 80%	Entities correctly cross-linked across document types in demo corpus
Time-to-Answer	< 5 seconds (P95)	End-to-end query against 10,000-document corpus
Compliance Gap Detection	> 88% precision	On known-gap test set from synthetic regulatory corpus

🎬 Demo Scenario

Narrative: A centrifugal pump (P-101A) at a petroleum refinery shows early vibration signatures.

KnowledgeBrain's response in under 90 seconds:

Copilot instantly surfaces the last 3 maintenance records, OEM vibration thresholds, and the last inspection report — in a single query with source citations.
Maintenance Agent cross-references similar past failures and recommends bearing inspection as the primary hypothesis, backed by two historical RCAs from 2021 and 2023.
Compliance Agent flags that an OISD-18 inspection is 47 days overdue and auto-drafts the compliance evidence package.
Lessons Learned Engine surfaces a 2019 incident where the same failure signature preceded a seal failure — and proactively alerts the shift supervisor.

The same workflow without KnowledgeBrain typically takes 2–4 hours of manual search across 7+ systems.

🗂️ Demo Dataset

All data is publicly available — no licensing issues:

Dataset	Source	Used For
OISD Incident Investigation Reports	oisd.gov.in (public)	Lessons Learned Engine
Sample P&IDs	ISA training resources	Computer vision pipeline
Factory Act text	India Code portal	Compliance corpus
PESO regulations	peso.gov.in	Compliance corpus
Synthetic work orders	`scripts/generate_demo_data.py`	Maintenance agent

Generate synthetic demo data:

python scripts/generate_demo_data.py --equipment-count 50 --years 5

🏆 Hackathon Context

Event: ET AI Hackathon 2026 Problem Statement: PS 8 — AI for Industrial Knowledge Intelligence Judging Criteria:

Criterion	Weight	Our Approach
Innovation	25%	Multi-modal knowledge graph across P&IDs, scanned forms, emails — proactive failure push
Business Impact	25%	35% of wasted hours recoverable; 18–22% downtime reduction; audit prep weeks → hours
Technical Excellence	20%	Multi-agent LangGraph; hybrid retrieval; CV P&ID parsing; confidence-scored answers
Scalability	15%	Container-native; Qdrant + Neo4j horizontally scalable; SaaS-ready for 100M+ docs
User Experience	15%	Mobile-first; sub-5s response; conversational; multi-language; offline-capable

👥 Team

Name	Role	Responsibilities
Saksham Sethi	Lead Architect	RAG pipeline, LlamaIndex integration, Qdrant vector store, API layer
Sumit Singh Charak	AI/ML Engineer	LLM prompt engineering, LangGraph agents, Maintenance & Compliance agents, evaluation
Gummadidala Sai Krishna	Full-stack & Infra	React/React Native UI, Neo4j knowledge graph, OCR/CV pipeline, deployment

🤝 Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you'd like to change. Please ensure tests pass before submitting:

pytest tests/ -v

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

📚 References

McKinsey Global Survey on Industrial Information Access, 2024
NASSCOM-EY Manufacturing & Energy Digital Maturity Study, 2024
BIS Research: Industrial Downtime Cost Analysis, Indian Heavy Industry, 2024
Anthropic API Documentation
LlamaIndex Documentation
Neo4j Knowledge Graph Docs

Built with ❤️ for ET AI Hackathon 2026 · PS 8

Making every piece of industrial knowledge queryable — instantly, accurately, at the moment it's needed.

Name		Name	Last commit message	Last commit date
knowledgebrain-api		knowledgebrain-api
.gitignore		.gitignore
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation