Document Semantic Search

Overview

An n8n workflow for building semantic search over your documents. Processes PDFs, docs, and text files, creates embeddings with OpenAI, stores them in Qdrant, and enables AI-powered search with Gemini.

How It Works

Ingestion Pipeline

1. Upload Documents - PDFs, DOCX, TXT, or URLs
2. Text Extraction - Convert to plain text
3. Chunking - Split into semantic chunks
4. Embeddings - Generate vectors with OpenAI
5. Storage - Store in Qdrant vector database

Search Pipeline

1. Query - Receive search query
2. Embed Query - Convert to vector
3. Vector Search - Find similar chunks in Qdrant
4. Context Assembly - Gather relevant chunks
5. AI Answer - Gemini synthesises response

Workflow Components

Features

• Multi-Format Support - PDF, DOCX, TXT, Markdown
• Smart Chunking - Preserves semantic boundaries
• Hybrid Search - Combines vector + keyword search
• Source Attribution - Links back to original documents
• Incremental Updates - Add new docs without full reindex

Architecture

Documents → Text Extraction → Chunking → OpenAI Embeddings
                                              ↓
                                          Qdrant DB
                                              ↓
Query → Embed Query → Vector Search → Gemini Answer

Example Query

User: “What’s our refund policy for software subscriptions?”

System:

Based on the company policies document:

**Software Subscription Refunds**

- Full refund available within 14 days of purchase
- Pro-rata refund for annual plans cancelled after 14 days
- No refund for monthly plans (cancel before renewal)

Special cases:
- Technical issues preventing use: Full refund at any time
- Billing errors: Immediate correction + refund

📄 Source: Company-Policies-2024.pdf (Page 12)

Use Cases

Knowledge Base - Search internal documentation

Legal/Compliance - Find relevant policy sections

Research - Search academic papers and reports

Customer Support - Find answers in product docs

Configuration