This repository contains a Streamlit-based chatbot designed to assist OT (Operational Technology) service and support teams by answering questions based on a library of Standard Operating Procedure (SOP) documents.
The system uses a RAG (Retrieval-Augmented Generation) pipeline built with Langflow to provide accurate, context-aware answers from your private documentation.
Watch the walkthrough on YouTube
- Conversational UI: An intuitive chat interface for asking technical questions.
- Document Upload: Easily upload and index new SOPs (PDF, DOCX, TXT).
- RAG Pipeline: Leverages a powerful backend to retrieve relevant document chunks and generate answers with an LLM.
- Session Management: Keeps track of conversation history and uploaded files per session.
-
Langflow Backend
- Install Langflow & MongoDB driver:
uv pip install langflow pymongo - Start the Langflow server:
uv run langflow run - Import the flow definitions from the
./Langflow/directory.
- Install Langflow & MongoDB driver:
-
Backend Service
- Navigate to the backend:
cd Backend - Install dependencies:
pip install -r requirements.txt - Run the FastAPI server:
uvicorn main:app --reload
- Navigate to the backend:
-
Frontend UI
- Navigate to the frontend:
cd Frontend - Install dependencies:
pip install -r requirements.txt - Run the Streamlit app:
streamlit run app.py
- Navigate to the frontend:
Whenever we make a significant architectural choice for this OT-Debug Chatbot, we record it here.
Date: 2025-06-16 Status: Accepted
---
config:
theme: redux-dark
layout: fixed
look: classic
---
flowchart TD
n5["Backend /upload"] --> n6["Langflow DataLoader Flow"]
n6 --> n7["Text Splitter"]
n7 --> n8["Chunks"]
n8 --> n9["MongoDb"]
n10["Ollama Embeddings"] --> n8
n12["Backend /query"] --> n13["Langflow RAG Flow"]
n13 --> n14["Chat input"]
n14 --> n16["MongoDb"]
n15["Ollama Embeddings"] --> n16
n17["Chunks"] --> n16
n16 --> n19["LLM for Generation"]
n2["User uploads SOP"] --> n5
n11["User chats"] --> n12
n5@{ shape: rect}
n6@{ shape: rect}
n9@{ shape: db}
n10@{ shape: rect}
n12@{ shape: rect}
n13@{ shape: rect}
n14@{ shape: rect}
n16@{ shape: db}
n15@{ shape: rect}
n17@{ shape: rect}
n19@{ shape: rect}
n2@{ shape: rect}
n11@{ shape: rect}
n5:::Sky
n5:::Rose
n5:::Aqua
n6:::Sky
n6:::Rose
n6:::Aqua
n6:::Pine
n7:::Sky
n8:::Sky
n10:::Sky
n12:::Sky
n12:::Rose
n12:::Aqua
n13:::Sky
n13:::Rose
n13:::Aqua
n13:::Pine
n14:::Sky
n15:::Sky
n17:::Sky
n19:::Sky
n2:::Sky
n2:::Rose
n11:::Sky
n11:::Rose
classDef Rose stroke-width:1px, stroke-dasharray:none, stroke:#FF5978, fill:#FFDFE5, color:#8E2236
classDef Aqua stroke-width:1px, stroke-dasharray:none, stroke:#46EDC8, fill:#DEFFF8, color:#378E7A
classDef Pine stroke-width:1px, stroke-dasharray:none, stroke:#254336, fill:#27654A, color:#FFFFFF
classDef Sky stroke-width:1px, stroke-dasharray:none, stroke:#374D7C, fill:#E2EBFF, color:#374D7C
We need a lightweight, interactive web UI that allows enterprise users to type in OT system issues and get back suggestions. It should be simple to deploy alongside our backend service.
Use Streamlit for the front end.
- Pro: Rapid prototyping, minimal boilerplate, built-in support for chat-style layouts.
- Con: Limited custom component flexibility (but acceptable for v1).
Date: 2025-06-16 Status: Accepted
We need a RESTful service to receive queries from the UI and forward requests to the Langflow server's run API.
Implement the backend in Python using FastAPI.
- Pro: Async support, automatic OpenAPI docs, easy integration with
httpx. - Con: Adds an extra hop between UI and Langflow (acceptable for decoupling).
Date: 2025-06-16 Status: Accepted
We have many SOP documents that must be searched and fed to an LLM dynamically. We want to keep document parsing, vector storage, and prompt templating in a single flow.
Utilize a standard Langflow flow, triggered via its API endpoint, that:
- Loads & splits SOPs in
./test_SOPs/ - Indexes them in a MongoDB Atlas collection specific to the session
- Retrieves top-k chunks at query time
- Passes them plus user query into an OpenAI LLM chain
- Pro: All RAG logic lives in one place, easy to modify prompts or retriever parameters.
- Con: Requires Langflow server uptime; introduces dependency on Langflow.
Date: 2025-06-17 Status: Accepted
We need a scalable, cloud-native vector store that supports session-specific data isolation for fast similarity searches over our SOP documents. The initial on-disk FAISS approach was not suitable for this.
Use MongoDB Atlas as the vector store. Each user session from Streamlit, identified by its session ID, will correspond to a unique collection. The vector search index will be configured to use cosine similarity.
- Pro: Fully managed service, scales on demand, and provides clear data isolation between user sessions.
- Con: Introduces a dependency on MongoDB Atlas and requires network access from the Langflow environment.
Date: 2025-06-16 Status: Accepted
We need a powerful model to interpret SOP context and troubleshoot OT-system issues effectively.
Use Gemini 2.5 Pro via Google AI.
- Pro: State-of-the-art reasoning, multimodality, and instruction-following.
- Con: Cost per token; dependency on Google Cloud services.
(When a new major decision arises—e.g. switching to Chroma, adding caching middleware, or upgrading the UI framework—append a new ADR with a fresh ID and date.)