1 min read 188 words Updated Sep 24, 2026 Created Sep 24, 2026
##NLP#Data#LLM#RAG#ml#review

Typical RAG pipelines usually include:

  • Preprocessing data basis, e. g. with #NLP
  • generating embeddings
  • similarity search on embeddings with given query
  • piping outputs from the vector DB, e. g. found similar documents into an LLM
    Thus, the LLM generated is augumented by the retrieval system.

Tools

Various tools exist that can be included into RAG pipelines. Here is an overview.

Memory

Hypothetical document embedding

See Hypothetical Document Embedding (HyDE)

Query rewriting / generation

A similar approach to Hypothetical Document Embedding (HyDE), but not an hypothetical answer is generated, but the query is enhanced by an LLM.

Architectures

Architectures of RAG pipelines are very close to th e

Self-retrieval

This architecture resolves around the model autonomously generating retrival queries. In most RAG pipelines, the embeddings and retrieval overall is justed based on user input. This approach might help with retrieval quality, as it lets the model identify what else might be needed.

See: Hybrid Search

Branched retrieval

Agentic RAG

Resources