Embedding Service

Local Testing UI

Embedding Service Architecture

CPU-deployable RAG pipeline with multilingual embeddings, hybrid search, and AI-powered metadata extraction

384
Embedding Dimensions
3
Search Modes
100+
Languages Supported

System Architecture

API Layer
Embeddings
POST /v1/embeddings
Search
POST /v1/search
PDF Upload
POST /v1/pdf-upload
Extract
POST /v1/extract
Processing Layer
ONNX Embedder
multilingual-e5-small
CPU-Optimized
BM25 Encoder
Sparse vectors
Keyword Search
Reranker
Cross-encoder
Relevance Boost
PDF Processor
Docling + Chunking
Document Parsing
LLM Extractor
OpenRouter API
Metadata AI
Storage Layer
Qdrant Vector Database
High-performance vector search engine
v1.16.2
Cosine Similarity
HNSW Index

Core Capabilities

Text Embeddings

Generate 384-dimensional semantic vectors for 100+ languages using E5-small model

Multilingual CPU-Fast ONNX

Hybrid Search

Combine semantic (dense) and keyword (sparse) search with Reciprocal Rank Fusion

Dense + Sparse RRF BM25

Cross-Encoder Reranking

Improve search relevance by reranking top results with cross-encoder model

Precision Boost Optional

PDF Processing

Automatic PDF parsing, chunking, and embedding using Docling document understanding

Docling Auto-Chunk Batch

LLM Extraction

Extract structured metadata from documents using AI-powered LLM analysis (Romanian + English)

OpenRouter Qwen Bilingual

Vector Search

Fast semantic search with metadata filtering, score thresholds, and pagination support

HNSW Filtering Scalable

Specs Management

View, search, and filter extracted specifications across all documents in a collection

Searchable Filterable Organized

Example Workflows

Real test data from

Using mock data - click refresh to run real benchmarks

1

Simple Semantic Search

Query: "How do I reset my password?"
POST /v1/embeddings β†’ Generate 384-dim vector
Search Qdrant with cosine similarity
POST /v1/collections/docs/search β†’ Top 5 results
2

Hybrid Search + Rerank

Dense (semantic) + Sparse (BM25)
Parallel search, merged with RRF
Fetch top 50 candidates
rerank_top_k: 50 β†’ Cross-encoder scores
Return top 5 reranked results
Precision-optimized ranking
Improved relevance:
3

PDF Ingestion with AI Extraction

Upload product_manual.pdf
POST /v1/collections/products/pdf-upload
LLM extracts metadata (Romanian)
Embed all chunks β†’ Store in Qdrant
47 vectors + metadata stored
4

Cross-Language Retrieval

Query in English: "warranty information"
E5 model understands semantic intent
Finds Romanian documents
Matches "garanΘ›ie 2 ani" in stored docs
Cross-lingual magic:

Quick API Examples

1

Generate Embeddings

POST /v1/embeddings
{
  "input": "Hello world",
  "input_type": "document"
}
β†’ [0.023, -0.145, ..., 0.091]
384 dimensions, ~50ms
2

Hybrid Search

POST /v1/collections/docs/search
{
  "query": "password reset",
  "search_mode": "hybrid",
  "rerank": true,
  "top_k": 5
}
β†’ 5 results, reranked
3

Upload PDFs

POST /v1/collections/docs/pdf-upload
FormData:
  files: manual.pdf
  chunk_size: 400
  extract_specs: true
β†’ 47 chunks, metadata extracted
4

Extract Metadata

POST /v1/extract
{
  "text": "Router TP-Link...",
  "language_hint": "ro"
}
β†’ product, specs, confidence

Technology Stack

🐍
Python 3.11+
Runtime
⚑
FastAPI
Web Framework
πŸ”’
ONNX Runtime
Inference
πŸ—„οΈ
Qdrant
Vector DB
πŸ“„
Docling
PDF Parser
πŸ€–
OpenRouter
LLM API
🐳
Docker
Deployment
πŸ“¦
uv
Package Manager

Configuration

Configure your API key and service endpoint

Default: test-api-key

The embedding service endpoint

Health Dashboard

Monitor service and dependency status

Service Status

Generate Embeddings

Convert text into 384-dimensional vector embeddings

Response


          

        

      


      
      

        

          

            

              
Collections

              
Manage your vector collections

            

            
          


          

            
New Collection

            

              

                
                
              

              

                
                
              

              

                
                
              

            

          


          

            

            
          

        

      


      
      

        

          
Similarity Search

          
Find similar vectors using text queries


          

            

              

                
                
              


              

                
                
              


              

                
                

                  
                  
                  
                

              


              

                
                
                

                  More Keyword
                  Balanced
                  More Semantic
                

              


              

                

                  
                  
                

                

                  
                  
                

              


              

                

                  
                  
                

                
Rerank results using a cross-encoder model for improved relevance


                

                  
                  
                  
Number of candidates to fetch before reranking (min: 10)

                

              


              
            

          


          

            

              
Search Results

              
              
                
                  
                
                Reranked ( candidates)
              
            

            
            

              
            

          

        

      


      
      

        

          
Extracted Specifications

          
View and search all extracted specifications from uploaded PDFs


          
          

            

              

                
                
              


              

                
                
              


              

                
                
              


              

                
                
              

            


            

              
              
            

          


          
          

            

            
Loading specifications...

          


          
          

            

          


          
          

            

              

                Found  documents with extracted specs
              

            


            

              
            

          


          
          

            
              
            
            
No specifications found

            

              
              
            

          

        

      


      
      

        

          
Metadata Extraction

          
Extract structured metadata from text using LLM (OpenRouter API)


          

            
            

              
              
              
Text will be truncated to 14,000 characters


              
              

                

                

                  
                  

                    
                    
                  


                  
                  

                    
                    
                    
Optional: Define custom fields to extract

                  

                

              


              
            


            
            

              

                
Extracted Metadata

                

                  

                    Confidence:
                    
                  

                  

                     fields
                  

                

              


              
              

                
                

                  
Product

                  

                


                

                  

                    
Manufacturer

                    

                  

                  

                    
Model Number

                    

                  

                


                
                

                  
Specifications

                  

                    
                  

                


                
                

                  
Categories

                  

                    
                  

                


                
                

                  

                    Language: 
                  

                  

                    Tokens used: 
                  

                


                
                

                  
                  

                    

                    
                  

                

              

            

          

        

      


      
      

        

          
PDF Upload

          
Upload PDFs to automatically chunk, embed, and store


          

            

              
              
            


            

              

                
                
              


              

                
              

            


            

              
Configuration

              

                

                  
                  
                


                
                

                  
                  

                    
                    
                  

                  
Extract product specs using AI (requires API key)

                

              


              
              

                

                

                  
                  
                  
Optional: Define custom extraction fields

                

              


              
            


            
            

              

                
                  
                  
                
                

                  
Processing PDFs...

                  
Converting  file(s), chunking with Docling, generating embeddings, and storing vectors

                

              

              

                

                  

                

              

            


            

              
Upload Results


              
              

                

                  

                  
Total PDFs

                

                

                  

                  
Successful

                

                

                  

                  
Chunks Created