The Problem
Organizations and individuals deal with large volumes of PDF documents — resumes, reports, manuals, contracts, and research papers. Extracting specific information from these documents is slow, manual, and error-prone.
Key pain points:
- Reading through entire PDFs to find one specific fact takes significant time
- Traditional search (Ctrl+F) only finds exact keyword matches, not semantic meaning
- No conversational interface — users cannot ask follow-up questions or get contextual answers
- Multiple PDFs cannot be queried together in a unified chat experience
- No session isolation — in multi-user environments, one user’s data bleeds into another’s
There was a need for an intelligent system that could read any PDF, understand its content semantically, and answer natural language questions — just like asking a human expert.
Our Solution
We built an AI-powered PDF Chatbot using Next.js, LangGraph, Qdrant vector database, and Groq LLM. The system allows users to:
- Upload any PDF document and chat with it instantly
- Ask questions in natural language — Hindi or English — and get accurate answers from the PDF content
- Upload multiple PDFs in one session and query across all of them
- Have normal conversations when no PDF is uploaded (general AI assistant mode)
- Start a new chat that clears only their own session data — other users’ data is unaffected
The solution uses semantic vector search — the chatbot understands the meaning of a question, not just keywords. It finds the most relevant chunks from the uploaded PDF and passes them to the LLM to generate accurate, contextual answers.
LangGraph orchestrates the entire pipeline as a stateful graph — routing, retrieval, and prompt construction happen as automatic sequential steps, making the system extensible and maintainable.
Solution Architecture
Upload Flow
When a user uploads a PDF:
- pdf-parse extracts raw text from the PDF
- RecursiveCharacterTextSplitter splits text into 500-character chunks with 50-character overlap
- HuggingFace (all-MiniLM-L6-v2) converts each chunk into a 384-dimensional vector
- Vectors are stored in Qdrant Cloud with sessionId and source (PDF filename) as metadata
Chat Flow — LangGraph Pipeline
Every user message goes through a 3-node LangGraph StateGraph:
| Node | Input (reads from state) | Output (writes to state) |
| routeNode | question, uploadedFiles, currentFile | isSimple = true/false — simple greeting hai to Qdrant search skip |
| retrieveNode | isSimple, question, sessionId, currentFile | context = Qdrant se relevant PDF chunks (score > 0.05 filter) |
| promptNode | context, uploadedFiles | systemPrompt = LLM ke liye sahi instruction (3 cases) |
State automatically travels node to node — routeNode sets isSimple, retrieveNode reads it and sets context, promptNode reads context and sets systemPrompt. No manual passing required.
3 Scenarios
Scenario 1 — Simple greeting (“hi”, “hey”, “thanks”):
- routeNode: isSimple = true
- retrieveNode: Qdrant search SKIP, context = “”
- promptNode: “Friendly AI assistant” prompt
- Result: Normal conversational reply
Scenario 2 — PDF attached + any message:
- routeNode: isSimple = false (currentFile set hai)
- retrieveNode: Sirf us PDF ke chunks search (source filter)
- promptNode: “Answer ONLY from this PDF content” prompt
- Result: Answer sirf us attached PDF se
Scenario 3 — Question without attaching PDF:
- routeNode: isSimple = false
- retrieveNode: Saare session ke uploaded PDFs mein search (sessionId filter only)
- promptNode: “Answer from all uploaded PDFs” prompt
- Result: Answer from any relevant uploaded PDF
Session Isolation
- Every user gets a unique sessionId stored in localStorage
- All Qdrant vectors tagged with sessionId — users never see each other’s data
- New Chat: /api/new-chat deletes only that user’s vectors, generates new sessionId
- Payload indexes on sessionId and source enable fast filtered queries
Deliverables
- Fully functional AI PDF Chatbot web application (Next.js 16)
- PDF upload and processing pipeline (pdf-parse + chunking + vector embedding)
- LangGraph-orchestrated 3-node pipeline (route → retrieve → prompt)
- Qdrant Cloud vector database integration with session-based isolation
- Real-time streaming chat interface with markdown rendering and code syntax highlighting
- PDF support — upload and query across multiple documents in one session
- New Chat functionality — session-scoped data cleanup
- Responsive UI (mobile + desktop) with file attachment, copy button, and streaming tokens
- REST APIs: /api/chat (streaming), /api/upload (PDF), /api/clear-context (session clear)
Tech Stack
| Technology | Purpose | Why chosen |
| Next.js 16 | Full-stack framework, API routes | React-based, App Router, SSE support |
| LangGraph | AI pipeline orchestration | Stateful graph, nodes auto-chain |
| LangChain Core | Message types, LLM interface | Standard AI message format |
| Groq API (Llama 3.3 70B) | LLM — answer generation | Free tier, fast, high quality |
| Qdrant Cloud | Vector database | Free tier, payload index, fast filter |
| HuggingFace API | Text embeddings (384-dim) | Free, all-MiniLM-L6-v2 model |
| pdf-parse | PDF text extraction | Simple, reliable, no binary deps |
| RecursiveCharacterTextSplitter | Text chunking | Smart overlap, LangChain built-in |
| ReactMarkdown + rehype | Markdown rendering in chat | Code blocks, tables, formatting |
| Prism Syntax Highlighter | Code highlighting in responses | 50+ languages, oneDark theme |
| TypeScript | Type safety across codebase | Prevents runtime errors |
| Tailwind CSS | UI styling | Utility-first, dark mode ready |
Business Impact
This solution has significant potential across multiple industries and use cases:
HR & Recruitment
- Recruiters can upload resumes and query across all of them — “which candidates know React?” — saving hours of manual screening
- Candidates can upload their own resume and ask the AI to help improve it or answer interview prep questions
Legal & Compliance
- Lawyers can upload contracts and ask specific questions — “what are the termination clauses?” — without reading 50 pages
- Compliance teams can query regulatory documents to find specific obligations instantly
Education & Research
- Students can upload research papers and textbooks and ask questions in their native language
- Professors can build course-specific chatbots by uploading study material
Enterprise Knowledge Management
- Employees can query internal SOPs, product manuals, and policy documents conversationally
- Onboarding time reduces significantly — new employees can get instant answers from documentation
Healthcare
- Doctors can upload patient reports and ask summarization or analysis questions
- Medical students can query textbooks and clinical guidelines in plain language
Key Metrics (Potential)
- 80% reduction in time to find specific information in documents
- Zero additional cost — entire stack runs on free tiers (Groq, HuggingFace, Qdrant)
- Multi-language support — answers in the same language as the question
- Infinitely scalable — Qdrant Cloud handles millions of vectors
