Web apps built to scale Since 2015
2MUCHCOFFEE Get a quote
Author

Oleg Logvin

AI development PDF Parsing for RAG: pdfplumber, PyMuPDF, Docling, Marker and MinerU on Page 300 of a Long Manual We ran pdfplumber, pypdf, PyMuPDF, pymupdf4llm, Docling, Marker and MinerU over a 400-page slice of the PostgreSQL manual, a two-column paper and a 1947 scan, and measured headers in the chunk, footnotes, a page-spanning table, OCR and cost per 1,000 pages, with the licence beside each. AI development PDF Chunking for RAG: Route OCR by Page, Keep Table Headers Attached, Re-index the Delta Every RAG demo works on a clean PDF. Then you point it at a decade of scanned contracts and it returns confident nonsense. The chunk is the unit of retrieval, and three moves decide whether yours are any good: route OCR by page, keep table headers attached, re-index the delta. AI development Agentic RAG: Why Retrieval Is the Layer Under the Agent Classic RAG is a straight line: retrieve once, then answer. Agentic RAG is a control loop, an agent that decides what to fetch, whether it was enough, and whether the question is answerable at all. Here is the real graph, and when the loop is worth its cost. AI development How to Actually Evaluate Your RAG: Evaluation Metrics, Golden Datasets, and the Regression Gate 'It feels better' is not a number. RAG evaluation metrics for both halves, a six-step golden dataset build, a pytest regression gate and two judge checks, so your next RAG change is measured before it ships. AI development Vector Database for RAG: Pinecone vs Qdrant vs pgvector (2026) The best vector database for RAG is the one that fits your hardest constraint. A Tech Lead who runs Qdrant compares Pinecone, Qdrant, pgvector, Weaviate, Milvus and Chroma on data residency, scale, hybrid search and filtered recall, every vendor fact checked against its docs on 30 October 2026. AI development Your Agent Is Only as Smart as Its Search: Inside Our Hybrid RAG on Qdrant Our matching agent answered a continuous stream of developer requests with confident, well-rounded, wrong shortlists, until we fixed the search under it. The architecture, the reasoning, and an interactive fusion demo you can tune yourself.