Skip to content

06 Adv Systems

TL;DR


graph LR

NR["Naïve RAG"]
AR["Advanced RAG"]
MR["Modular RAG"]

NR --> AR --> MR

NR -->|Process| P1["Retrieve ⟶ Read"]

AR -->|Enhancements| P2["Pre-R ⟶ Retrieval ⟶ Post-R"]
P2 -->|Goal| G1["Higher Precision & Recall"]

MR -->|Architecture| P3["Interchangeable Modules"]
P3 -->|Goal| G2["Flexibility & Scalability"]

RAG System

  • subtypes ⟶ {Naïve, Advanced, Modular}

  • pipelines ⟶ {Indexing, Generation}

Sam

#1. Naïve

  • implements_framework: Retrieve ⟶ Read

#2. Advanced

  • extends: Naïve

  • implements_framework: Rewrite ⟶ Retrieve ⟶ Re-rank ⟶ Read

  • composed_of_stages: {Pre-R, Retrieval, Post-R}

  • aims_to ⟶ improve precision, recall, and contextual alignment

#3. Modular

  • extends ⟶ Advanced RAG

  • decomposes_into_modules ⟶ {Core, New}

1. Naïve

Limitations

  • R ⟶ {low precision, low recall}

  • A ⟶ {redundancy, disjoint context, context length limits}

  • G ⟶ {hallucination, bias, over-reliance on retrieved context}

2. Advanced

2.1 Pre-R Stage

  • Index Optimization: Optimize our KB.

  • Query Optimization: Optimize our user Q before retrieval.

Sam

Index Optimization

  • Chunk: Chunk Size Tuning · Context-Enriched Chunking · Surrounding-Chunk Retrieval

  • Metadata: Metadata Filtering · Metadata Enrichment

  • Index: Parent-Child Hierarchy · Knowledge Graph Index

Sam

Query Optimization

  • Query Expansion: Multi-Query · Sub-Query · Step-Back

  • Query Transformation: Rewrite · HyDE

  • Query Routing: Intent-Based · Metadata-Based · Semantic-Based

2.2 Retrieval Stage

Sam

  • uses_strategies:

    • Hybrid R ⟶ combines sparse + dense + graph retrieval

    • Iterative R ⟶ loops retrieval using generated outputs

    • Recursive R ⟶ transforms query iteratively

    • Adaptive R ⟶ employs LLMs to decide when/what to retrieve

    • is_a_subtype_of ⟶ Agentic AI

2.3 Post-R Stage

Sam

  • includes:

    • Compression ⟶ removes irrelevant tokens, fits LLM context window

    • Re-ranking ⟶ prioritizes retrieved docs for generation

3. Modular

  • extends ⟶ Advanced RAG

  • decomposes_into_modules ⟶ {Core, New}

3.1 Core

Sam

Modules:

  • I ⟶ builds KB, manages embeddings & chunking

  • R ⟶ enables interchangeable retrievers

  • G ⟶ manages LLM selection & prompt augmentation

  • Pre-R ⟶ encapsulates Pre-R techniques

  • Post-R ⟶ encapsulates Post-R techniques

3.2 New

Sam

  • Search ⟶ expands access to multiple data sources

  • Fusion ⟶ aggregates multi-query results

  • Memory ⟶ leverages LLM parametric memory

  • Routing ⟶ directs queries through optimal paths

  • Task Adapter ⟶ adapts system for specific downstream tasks