AI & retrieval systems
Doc Chat
Route the question. Retrieve the evidence. Then ask the model.
Doc Chat is a retrieval system before it is a chat interface. A React client talks to a Python backend that selects the right retrieval operation, gathers evidence, and only then uses an LLM to compose an answer.
Open public projectSYSTEM ARCHITECTURE
Different questions take different retrieval paths
A query router can choose structured PostgreSQL lookup, semantic retrieval through embeddings and pgvector, or a hybrid of both. Candidates are reranked before the best evidence reaches the language model.
ENGINEERING STORY
Documents become retrievable evidence
Ingestion separates parsing, chunking, metadata, and embedding creation. The result supports both semantic similarity and deterministic filtering over the same document collection.
Not every question belongs in vector search
Questions about identifiers, ownership, counts, or metadata can use structured queries. Content questions use semantic retrieval. Mixed questions combine both sources before answer generation.
Probabilistic reasoning has deterministic boundaries
An agent-style workflow selects focused operations such as entity resolution, structured querying, semantic search, hybrid retrieval, and answer generation. The model is part of the control flow, not a replacement for application logic.
Reranking improves what the model gets to see
Initial retrieval produces candidate passages. A reranking stage prioritizes the strongest evidence so the final generation step works with a smaller, more relevant context.
Designed for Hungarian and English collections
The document and question-processing workflow accounts for both Hungarian and English content without claiming language support beyond those implemented use cases.
TECHNICAL FOUNDATION
Layers of the system
Frontend
AI backend
Retrieval data
Implemented capabilities
- Document ingestion, parsing, chunking, and embedding pipelines
- Semantic, structured, and hybrid retrieval paths
- Tool-oriented query routing and entity resolution
- Candidate reranking before LLM-based answer generation
- Hungarian and English document/question processing