How AI Transforms COI Management: A Real Factory Use Case (Hybrid Rasa + LangChain)
Managing COIs (Certificates of Inspection) inside a factory is usually a slow and manual process. QC staff search for customers, check lot numbers, look up QC results, generate Excel files, and manually send reports.
At Simplico, we built a Hybrid AI Assistant powered by Rasa (structured conversational control) + LangChain (agentic reasoning + SQL tools + RAG) to automate these workflows using a simple chat interface.
In this article, we walk through a real end-to-end use case from our Factory Automation POC and show the full system architecture diagram behind it.
🧠 System Architecture (Hybrid Rasa + LangChain)
Before jumping into the story, here is the architecture that powers the COI AI assistant:
┌─────────────────────────────┐
│ Factory Staff │
│ (QC, Engineer, Sales, Admin) │
└──────────────┬───────────────┘
│
▼
┌─────────────────────────────┐
│ Web Chat UI │
│ (Browser, LINE, Slack, etc.)│
└──────────────┬───────────────┘
│
▼
┌────────────────────────────────────┐
│ AI Gateway Layer │
│ Routes messages to Rasa/LangChain │
└───────┬────────────────────────────┘
│
┌──────────────────┼───────────────────┐
│ │ │
▼ ▼ ▼
┌────────────────┐ ┌───────────────────┐ ┌───────────────────────┐
│ Rasa NLU + │ │ LangChain RAG │ │ LangChain Agent │
│ Dialogue Core │ │ Document Search │ │ Tools + Reasoning │
└────────────────┘ │ (Manual, SOPs) │ └───────────────────────┘
│ └───────────────────┘ │
│ │
▼ ▼
┌──────────────────────────┐ ┌────────────────────────────────┐
│ Rasa Custom Actions │ │ Agent Toolsets │
│ (Python logic, routing) │ │ - SQL Query Tool │
│ - Entity disambiguation │ │ - COI XLSX Generator Tool │
│ - User selection logic │ │ - Lot Lookup Tool │
└───────────────┬──────────┘ │ - Product Data Tool │
│ └───────────────┬────────────────┘
▼ ▼
┌────────────────────────────────┐ ┌──────────────────────────────────┐
│ COI / QC Database │ │ Document & Vector Store │
│ (Customers, Products, Lots, │ │ (Manuals, SOPs, Embeddings) │
│ Test Results, History, etc.) │ │ FAISS / PGVector │
└────────────────────────────────┘ └──────────────────────────────────┘This hybrid model combines the safety of structured dialog with the intelligence of agentic AI—making it perfect for real factory environments.
🔍 1. Searching Customer Data with Natural Conversation
Instead of browsing menus, the staff simply types:
User:
“Show me data of Toyota”
The AI instantly looks up the COI database.
Since factories often have multiple similar customers, the assistant clarifies:
AI:
“I found multiple customers named ‘Toyota’.
Please choose one:
- Toyota Motor
- Toyota Thailand
- Toyota Industrial Asia
- Toyota Logistics Service”
This disambiguation is generated by LangChain SQL tools but controlled safely by Rasa dialogue rules.
The user chooses:
User:
“2”
AI:
“You selected Toyota Thailand.
Here is the customer profile:
- Customer Code: C-TH-112
- Total COIs issued: 243
- Active product codes: 18
- Address: Samut Prakan, Thailand”
🧾 2. Generating a COI File
Now the user requests a COI.
User:
“Create COI for lot 123-456A”
AI performs:
- Lot lookup
- QC result check
- Template loading
- XLSX generation
- Upload to COI storage
AI:
“✔ Lot found
✔ QC: Passed
✔ Using template: COI_Template_v3.xlsx
Generating COI…”
After 2 seconds:
AI:
“Your COI is ready:
📄 Toyota_Thailand_COI_123-456A.xlsx
Download: s3://coi/Toyota_Thailand_COI_123-456A.xlsx”
The entire COI workflow—normally 5–10 minutes—finishes instantly.
📦 3. Listing Lots Under a Product Code
User:
“List lot numbers for product code P-9981”
AI:
“Here are the lots for P-9981:
- 123-456A — Completed
- 123-456B — Awaiting QC
- 123-457A — Completed
- 123-458C — Testing
- 123-460A — Completed (Pending COI)
Would you like to see any lot?”
User:
“Okay”
AI:
“Please choose 1–5.”
🏭 4. Inspecting Lot Data
User:
“1”
The AI returns a full QC + production summary:
Product
- Code: P-9981
- Name: Aluminum Housing
QC Results
- Hardness: 88.5 HRC
- Chemical Composition: OK
- Dimension: Passed
- Visual: No defects
Production
- Machine: CNC Line #2
- Date: 2025-01-14
- Operator Team: B
- Cycle Time: 2.8 minutes
Then the AI asks:
“Would you like to generate the COI, export to Excel, or compare with other lots?”
📘 5. Retrieving SOPs Through RAG
User:
“Show SOP for this product.”
AI (via RAG search):
“Here are the SOP steps for P-9981:
- Visual inspection
- Measure dimensions (±0.05 mm)
- Hardness test (85–90 HRC)
- XRF analysis
- COI preparation”
Everything comes directly from the factory’s SOP documents using vector embeddings + LangChain.
📊 6. Summarizing COI History
User:
“Summarize COI history for Toyota Thailand.”
AI:
COI Summary – Last 90 Days
- Total COIs: 62
- QC Pass Rate: 98.3%
- Delayed COIs: 2
- Main product: P-9981
“Would you like this as PDF, Excel, or email?”
🚀 Final Thoughts
This use case shows how a Hybrid Rasa + LangChain AI Assistant can transform COI workflows:
- Natural conversation instead of complex UI
- Automatic COI generation
- Fast lot lookup
- Instant QC insights
- One-click SOP retrieval
- Summary reports on demand
It’s not just a chatbot—it's an intelligent bridge connecting people, data, QC history, manuals, and COI documents into one seamless experience.
This POC proves that AI can save time, reduce errors, and unlock new productivity for factories of any size.
Get in Touch with us
Chat with Us on LINE
iiitum1984
Related Posts
- 基于启发式与新闻情绪的短期价格方向评估（Python）
- Estimating Short-Term Price Direction with Heuristics and News Sentiment (Python)
- Rust vs Python：AI 与大型系统时代的编程语言选择
- Rust vs Python: Choosing the Right Tool in the AI & Systems Era
- How Software Technology Can Help Chanthaburi Farmers Regain Control of Fruit Prices
- AI 如何帮助发现金融机会
- How AI Helps Predict Financial Opportunities
- 在 React Native 与移动应用中使用 ONNX 模型的方法
- How to Use an ONNX Model in React Native (and Other Mobile App Frameworks)
- 叶片病害检测算法如何工作：从相机到决策
- How Leaf Disease Detection Algorithms Work: From Camera to Decision
- Smart Farming Lite：不依赖传感器的实用型数字农业
- Smart Farming Lite: Practical Digital Agriculture Without Sensors
- 为什么定制化MES更适合中国工厂
- Why Custom-Made MES Wins Where Ready-Made Systems Fail
- How to Build a Thailand-Specific Election Simulation
- When AI Replaces Search: How Content Creators Survive (and Win)
- 面向中国市场的再生资源金属价格预测（不投机、重决策）
- How to Predict Metal Prices for Recycling Businesses (Without Becoming a Trader)
- Smart Durian Farming with Minimum Cost (Thailand)
