Types

25 of RAG

@Bhavishya Pandit
1. Standard RAG
Combines retrieval with large
language models for accurate,
context-aware responses.
Breaks documents into chunks for
efficient information retrieval.
Aims for 1-2 second response times
for real-time use.
Enhances answer quality by
leveraging external data sources.

2. Corrective RAG
Focuses on identifying and fixing
errors in generated responses.
Uses multiple passes to improve
outputs based on feedback.
Aims for higher precision and
user satisfaction compared to
standard RAG.
Leverages user feedback to
enhance the correction process .

@Bhavishya Pandit
3. Speculative RAG
Uses a small specialist model for
drafting and a larger generalist
model for verification, ensuring
efficiency and accuracy.
Parallel Drafting: Speeds up
responses by generating multiple
drafts simultaneously.
Superior Accuracy: Outperforms
standard RAG systems.
Efficient Processing: Offloads
complex tasks to specialized
models, reducing computational
load.

4. Fusion RAG
Integrates multiple retrieval
methods and data sources for
enhanced response quality.
Provides comprehensive answers
by leveraging diverse data inputs.
Increases system resilience by
reducing dependence on a single
source.
Adapts retrieval strategies
dynamically based on query
context.
3

@Bhavishya Pandit
 5. Agentic RAG
Uses adaptive agents for real-
time strategy adjustments in
information retrieval.
Accurately interprets user intent
for relevant, trustworthy
responses.
Modular design enables easy
integration of new data sources
and features.
Enhances parallel processing and
performance on complex tasks by
running agents concurrently.

6. Self RAG
Uses the model's own outputs as
retrieval candidates for better
contextual relevance.
Refines responses iteratively,
improving consistency and
coherence.
Grounds responses in prior
outputs for increased accuracy.
Adapts retrieval strategies based
on the conversation's evolving
context.
4

@Bhavishya Pandit
7. Adaptive RAG It dynamically decides when to
retrieve external knowledge,
balancing internal and external
knowledge.
It uses confidence scores from the
language model's internal states to
assess retrieval necessity.
An honesty probe helps the model
avoid hallucinations by aligning its
output with its actual knowledge.
It reduces unnecessary retrievals,
improving both efficiency and
response accuracy​.

REFEED refines model outputs

using retrieval feedback without
8. REFEED Retrieval Feedback
fine-tuning​.
Initial answers are improved by
retrieving relevant documents
and adjusting the response based
on the new information.
Generates multiple answers to
improve retrieval accuracy.
Combines pre- and post-retrieval
outputs using a ranking system to
enhance answer reliability​. 5

@Bhavishya Pandit
 9. REALM
REALM retrieves relevant documents
from large corpora like Wikipedia to
enhance model predictions​.
The retriever is trained with masked
language modeling, optimizing retrieval to
improve prediction accuracy​.
It uses Maximum Inner Product Search to
efficiently find relevant documents from
millions of candidates during training​.
REALM outperforms previous models in
Open-domain Question Answering by
integrating external knowledge​.

10. RAPTOR- Tree-Organized Retrieval

RAPTOR builds a hierarchical tree by
clustering and summarizing text
recursively​.
It enables retrieval at different
abstraction levels, combining broad
themes with specific details​.
RAPTOR outperforms traditional
methods in complex question-
answering tasks​.
Offers tree traversal and collapsed
tree methods for efficient
information retrieval​. 6

@Bhavishya Pandit
11. REVEAL for Visual-Language Model
This technique combines reasoning
with task-specific actions and external
knowledge, improving decision-
making.
It minimizes errors by grounding
reasoning in real-world facts, reducing
inaccuracies and hallucinations.
The method offers clear, human-like
task-solving steps, enhancing
transparency and interpretability.
REVEAL achieves strong performance
across tasks with fewer training
examples, making models efficient,
adaptable, and responsive.

12. REACT
The ReAct technique combines reasoning
and action, allowing models to interact
with their environment.
It maintains situational awareness by
updating context with past actions and
thoughts.
The model generates task-aligned
thoughts to guide logical decision-making.
Real-time feedback refines
understanding, reducing errors and
enhancing transparency and reliability. 7

@Bhavishya Pandit
13. REPLUG Retrieval Plugin REPLUG enhances LLMs by
retrieving relevant external
documents to improve prediction
accuracy.
It treats the language model as a
fixed "black box", prepending
retrieved information to the input.
This flexible design works with
existing models without
modifications, integrating external
knowledge to reduce errors and
hallucinations.
The retrieval component can be
Memo RAG combines memory and fine-tuned with model feedback,
retrieval to handle complex aligning better with the model’s
queries. needs and expanding niche
A memory model generates draft knowledge.
answers that guide the search for
external information.
The retriever then gathers relevant
data from databases, which a more 14. MEMO RAG
powerful language model uses to
create a comprehensive final
answer.
This method helps Memo RAG
manage ambiguous queries and
efficiently process large amounts
of information across various
tasks.
8

@Bhavishya Pandit
15. Attention-based RAG
ATLAS improves language models by
retrieving external documents to
enhance accuracy, especially in
question-answering tasks.
It uses a dual-encoder retriever to
identify the top-K relevant
documents from large text corpora.
A Fusion-in-Decoder model
integrates query and document
information, generating accurate
responses while reducing reliance on
memorization.
The document index is updatable
16. RETRO without retraining, ensuring it
RETRO splits input text into chunks and remains current and effective for
retrieves similar information from a knowledge-intensive tasks.
large text database to enrich context.
It uses pre-trained BERT embeddings to
pull in relevant chunks from external
data, enhancing context.
Chunked cross-attention integrates
these chunks, improving predictions
without a major increase in model size.
This approach enhances tasks like
question answering and text generation
efficiently, accessing extensive
knowledge with lower computational
demands than larger models. 9

@Bhavishya Pandit
17. AUTO RAG AutoRAG automates
optimization for Retrieval-
Augmented Generation (RAG)
systems.
It evaluates modules like
query expansion, retrieval,
and reranking for best
performance.
The framework uses a
modular, node-based
structure to test various
configurations.
18. CORAG :Cost- A greedy optimization
approach enhances
Constrained RAG efficiency across different
It enhances RAG by optimizing datasets.
relevant chunk selection from
databases.
It tackles three challenges:
correlating chunks efficiently,
handling non-monotonic utility
where adding chunks may reduce
utility, and adapting to diverse
query types.
CORAG uses Monte Carlo Tree
Search (MCTS) for optimal chunk
combination while factoring in
cost constraints, achieving up to
a 30% improvement over
baseline models. 10

@Bhavishya Pandit
 19. EACO-RAG
EACO-RAG enhances RAG with
edge computing for faster,
efficient responses.
Vector datasets are
distributed across edge
nodes, reducing delays and
resource use.
Adaptive knowledge updates
and inter-node collaboration
improve response accuracy.
20. RULE RAG A multi-armed bandit
approach optimizes cost,
Rule-RAG enhances question accuracy, and delay in real-
answering by adding rule-based time.
guidance to RAG.
It retrieves documents logically
relevant to queries using
predefined rules.
Rules are also used to guide
answer generation for accuracy
and context.
It includes in-context learning
(ICL) and a fine-tuned version
(FT) for better retrieval and
generation.

11

@Bhavishya Pandit
 21. Conversational RAG
CORAL benchmarks multi-turn
conversational RAG using
Wikipedia data.
It evaluates passage retrieval,
response generation, and
citation labeling.
CORAL handles open-domain,
realistic, multi-turn
conversations.
It bridges single-turn RAG
research and real-world multi-
turn needs.

22. Iterative RAG

Unlike traditional retrieval, iterative
RAG performs multiple retrieval steps,
refining its search based on feedback
from previously selected documents​.
Retrieval decisions follow a Markov
decision process​.
Reinforcement learning improves
retrieval performance.
The iterative retriever maintains an
internal state, allowing it to adjust
future retrieval steps based on the
accumulated knowledge from previous
12
iterations.

@Bhavishya Pandit
23. Context-driven Tree-structured It is a context-driven, tree-
structured RAG approach that
Retrieval decomposes complex queries into
hierarchical sub-queries,
enhancing retrieval depth.
Its workflow has two stages: a top-
down exploration of query facets,
creating a tree of retrieved
passages, followed by bottom-up
synthesis, integrating summarized
information to produce a coherent
long-form response.
This framework reduces gaps in
information and improves the
quality of generated content.

24. Causality-Enhanced Reflective and

Multi-Agent Framework: CRAT
enhances translation by detecting,
Retrieval-Augmented Translation
clarifying, and translating ambiguous
terms.
Knowledge Graph : Combines internal
and external sources to capture
context for accurate term use.
Causality Validation: A judge agent
validates information to ensure
context-aligned translations.
Refined Output: CRAT delivers precise,
consistent translations by using 13
validated knowledge.
@Bhavishya Pandit
 25. Graph RAG

Graph RAG constructs a knowledge

graph on-the-fly, linking relevant
entities during retrieval.
It leverages node relationships to
decide when and how much external
knowledge to retrieve.
Confidence scores from the graph
guide expansion, avoiding irrelevant
additions.
This approach improves efficiency
and response accuracy by keeping
the knowledge graph compact and
relevant.

14

@Bhavishya Pandit
 Bhavishya Pandit

Save Like Repost

15