Polski
Backend API · Web Interface Django · FastAPI · RAG

Semantic Search Engine

From your own documents to search and AI answers

GitHub — engine & API GitHub — user interface
  • Semantic Search
  • Hybrid Search
  • RAG
  • Milvus
  • PostgreSQL
  • Knowledge Base

What Is Semantic Search Engine?

Semantic Search Engine is a toolkit for building a knowledge base from your own documents. It prepares and indexes content, retrieves passages by the meaning of a question, and generates answers using the retrieved material through RAG (Retrieval-Augmented Generation).

The product has two cooperating parts: an engine with a REST API and a web interface. Use the ready-made panel or integrate the backend into your own application; the engine repository also includes the sse_lib Python client.

What Can You Use It For?

  • Searching documentation and internal material: Find related passages even when a question uses different wording from the document.
  • Team knowledge bases: Organize content into collections, browse documents, and decide which ones participate in search.
  • Source-based assistants: Build answers from retrieved passages rather than relying solely on a model's prior knowledge.
  • Search and RAG experiments: Choose embedding models, rerankers, filters, and generation settings; revisit query history and collect answer ratings.

Key Features

  • File uploads or pasted text: The interface accepts PDF, DOCX, TXT, HTML, Markdown, and ZIP archives. Content is prepared for indexing and split into chunks.
  • Chunking controls: Set chunk size and overlap for a collection. The panel takes the chosen embedding model's length limit into account.
  • Knowledge base management: Create and delete collections, inspect document content, delete chunks, and include or exclude documents from search results.
  • Indexing progress: Track processed, pending, and failed documents. The backend can index within the request or optionally through a Celery queue with RabbitMQ.
  • Semantic and hybrid search: Combine Milvus vector retrieval with PostgreSQL full-text search using RRF (Reciprocal Rank Fusion). An optional reranker reorders results against the question.
  • Filters and models: Narrow searches by categories and documents, among other criteria, and use configured embedding, reranking, and generative models.
  • RAG answers and feedback: Browse retrieval results first, followed by the model's answer loaded asynchronously. Rate answers with stars and comments, and revisit saved searches within a collection.
  • Polish and English interface: Switch languages and choose a light, dark, or automatic theme. Interface assets are served locally, without external CDN dependencies.

Architecture and Technology 

Browser → FastAPI + Jinja2 interface → Django REST API
                                          ├─ PostgreSQL: documents, metadata, and text search
                                          ├─ Milvus: vector index
                                          ├─ Embedding models and rerankers
                                          └─ Generative models: RAG answers

The backend uses Python, Django, and Django REST Framework. PostgreSQL stores relational data, while Milvus provides vector retrieval. Embedding models and rerankers match content to queries; a configured generative model creates an answer from the results.

The interface is a separate Python application built with FastAPI and Jinja2 templates, with local CSS and JavaScript. It communicates with the engine through the REST API and supports login using a backend token. It does not require a SPA framework.

The system can run on your own infrastructure. Communication with external services depends on the configured models and integrations: a local interface does not automatically mean all data processing is local. Retrieval relevance and answer quality depend on the documents, chunking, and chosen models; AI answers still require verification.

Code and Documentation