Skip to content

RAG Pipeline Recipes: From Your Parser to Your Vector Database ​

Every production RAG pipeline has the same three-part anatomy: a parsing/OCR step that turns documents into text, a vector database that stores and searches embeddings, and — between them — the step that decides retrieval quality: chunking.

The first two are usually already chosen. You run Mistral OCR because it fits your cost profile, or self-host PaddleOCR-VL for data locality; you run Qdrant because you like hybrid search, or pgvector because you already operate Postgres. Those are good decisions. The 104 recipes below keep both of them and fix the middle: POMA's bring-your-own-OCR connector consumes your tool's raw output, rebuilds document hierarchy, and emits chunksets shaped for your database's retrieval primitives — via vektoria's assemble() for the databases without a first-party POMA connector yet.

Browse by parser ​

Browse by database ​

The full matrix ​

Every cell links to a complete recipe: what the parser's raw output contains, how POMA's connector turns it into chunksets, a metadata-mapping table for that specific database, end-to-end code, and the failure mode you hit wiring the two together naively.

New to the anatomy? Start with the RAG chunking guide, the document ingestion guide, or the RAG architecture guide.