TotalApp Docs
Ctrl K
Home
EN | TR
Go to App

Knowledge / AI

Document loaders, text splitters, embeddings, vector databases, and rerankers for RAG pipelines.

Overview

The Knowledge / AI group provides the building blocks for retrieval-augmented generation (RAG) pipelines — loading documents, splitting them into chunks, embedding those chunks, storing them in a vector database, and reranking retrieved results before they reach an AI Agent node.

Nodes in This Group

NodeDescription
Document LoaderLoads a document (PDF, text, etc.) into the pipeline as raw text ready for splitting.
Text SplitterBreaks upstream text into smaller chunks sized for embedding.
Embeddings ModelConverts upstream text chunks into vector embeddings.
Vector DatabaseStores or queries embeddings in a vector database for similarity search.
RerankerReorders a set of retrieved candidates by relevance to a query before they are used downstream.

Usage Tips

1. Document Loader
→
2. Text Splitter
→
3. Embeddings Model
→
4. Vector Database

This is the typical ingestion chain for building a knowledge base: load a document, split it into chunks, embed each chunk, then store the embeddings. At query time, embed the incoming question, search the Vector Database node, optionally pass results through Reranker, then feed the top results into an AI Agent node as context.

Chunk size matters

Text Splitter's chunk size affects both retrieval quality and embedding cost. Smaller chunks give more precise retrieval but require more embeddings; larger chunks retain more context per chunk but can dilute relevance.

Frequently Asked Questions

Is Reranker required in a RAG pipeline?
No, it's optional. Reranker improves relevance ordering when the Vector Database returns many candidates, but a simple pipeline can skip it and pass Vector Database results directly to an AI Agent node.
Can I reuse a Vector Database node's stored embeddings across multiple workflows?
Yes — a Vector Database node connects to a persistent vector store. Any workflow that references the same store/collection can query embeddings written by a different workflow.