Retrieval-Augmented Generation (RAG)

Grounding generation in retrieved documents.

Tier Grades

Tier A: Frontier AI labs

feasible

medium confidence

NVIDIA ships Nemotron Retriever models for RAG.

Tier B: Indonesian AI labs

feasible

low confidence

RAG is training-light and weeks per system; the notes give no compute anchor.

Tier C: Institution units such as ITB

feasible

medium confidence

Academia is observed, and the method needs inference rather than training.

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

feasible

medium confidence

Integrators are observed and a system takes weeks, which fits proof-of-concept delivery; the notes give no compute anchor.

Where this comes from

Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

ComputeInference-heavy, training-light; compute not sourced.
DataClient or domain document collections; not quantified.
Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): not sourced in the notes.
Time horizonWeeks for a system.
Shelf-life signalsnot in the notes
Pipeline roleTier A supplies retriever and generator models; Tier C studies methods; Tier B and Tier D build grounded systems on local documents for proofs of concept and value.

Where this comes from

Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars

RAG survey (Naive, Advanced, Modular RAG) NVIDIA

Nemotron Retriever models

Observed actors

Academia

NVIDIAIntegrators

Linked models

Nemotron Retriever modelsembeddingexemplar

EmbeddingGemma 2embeddingused by

SEA-LION-E5-Embedding-600Membeddingused by

SEA-LION ModernBERT-Embedding 300Membeddingused by

SEA-LION ModernBERT-Embedding 600Membeddingused by

jina-reranker-v3.5rerankerused by

Nemotron-3-Embed-1Bembeddingused by

Nemotron-3-Embed-8Bembeddingused by

R3-rerank-0.6brerankerused by

Granite Embedding Multilingual R2embeddingused by

jina-embeddings-v5-omni-nanoembeddingused by

jina-embeddings-v5-omni-smallembeddingused by

Qwen3-VL-Embedding-2Bembeddingused by

Qwen3-VL-Embedding-8Bembeddingused by

BGE-VL familyembeddingused by

harrier-oss-v1-0.6bembeddingused by

jina-embeddings-v5-text-nanoembeddingused by

jina-embeddings-v5-text-smallembeddingused by

Qwen3-VL-Rerankerrerankerused by

Qwen3-VL-Reranker-2Brerankerused by

Qwen3-VL-Reranker-8Brerankerused by

ColEmbed V2embeddingused by

Llama Nemotron Embed VL 1B v2embeddingused by

Llama Nemotron Rerank VL 1B v2rerankerused by

+106

Llama Nemotron Embed 1B v2embeddingused by

Llama Nemotron Rerank 1B v2rerankerused by

bge-reasoner-embed-qwen3-8bembeddingused by

Granite Embedding Reranker English R2rerankerused by

EmbeddingGemma 300Membeddingused by

Granite Embedding small English R2embeddingused by

Qwen3-Embedding-0.6Bembeddingused by

Qwen3-Embedding-4Bembeddingused by

Qwen3-Embedding-8Bembeddingused by

Qwen3-Reranker-0.6Brerankerused by

Qwen3-Reranker-4Brerankerused by

Qwen3-Reranker-8Brerankerused by

jina-embeddings-v4embeddingused by

Gemini Embeddingembeddingused by

gemini-embedding-001embeddingused by

Sonar ProLLMproduced by

Sonar Reasoning Proreasoningproduced by

mxbai-rerank-large-v2rerankerused by

gte-reranker-modernbert-basererankerused by

nomic-embed-multimodalembeddingused by

Nomic embed text v2 MoEembeddingused by

Granite Embeddingembeddingused by

Granite Embedding 278M multilingualembeddingused by

ColQwen2 v1.0embeddingused by

jina-embeddings-v3embeddingused by

NV-Embed-v2embeddingused by

DataGemma RIG 27B ITLLMproduced by

DataGemmaLLMproduced by

DataGemma RAG 27B ITLLMproduced by

ColPali v1.2embeddingused by

gte-multilingual-baseembeddingused by

gte-multilingual-reranker-basererankerused by

ColPaliembeddingused by

jina-reranker-v2-base-multilingualrerankerused by

gte-Qwen2-7B-instructembeddingused by

gte-large-en-v1.5embeddingused by

snowflake-arctic-embed-lembeddingused by

Snowflake Arctic Embedembeddingused by

snowflake-arctic-embed-membeddingused by

Command R+ (Apr '24)LLMproduced by

Command R (Mar '24)LLMproduced by

mxbai-embed-large-v1embeddingused by

nomic-embed-text-v1.5embeddingused by

nomic-embed-text-v1embeddingused by

Nomic Embed Text v1embeddingused by

SFR-Embedding-Mistralembeddingused by

bge-m3embeddingused by

e5-mistral-7bembeddingused by

multilingual-e5-large-instructembeddingused by

MedCPT Cross-Encoderrerankerused by

jina-embeddings-v2-base-enembeddingused by

bge-large-en-v1.5embeddingused by

bge-large-zh-v1.5embeddingused by

bge-reranker-largererankerused by

bge-small-en-v1.5embeddingused by

bge-base-en-v1.5embeddingused by

BGE Reranker (v1)rerankerused by

bge-reranker-basererankerused by

BGE v1.5embeddingused by

GTE (General Text Embeddings)embeddingused by

gte-largeembeddingused by

gte-smallembeddingused by

multilingual-e5-smallembeddingused by

all-MiniLM-L6-v2 (Xenova, Transformers.js ONNX)embeddingused by

SPECTER2embeddingused by

SPLADE (CoCondenser self-distil)embeddingused by

paraphrase-multilingual-MiniLM-L12-v2embeddingused by

paraphrase-multilingual-mpnet-base-v2embeddingused by

GTR (Generalizable T5 Retriever)embeddingused by

Sentence-T5embeddingused by

SapBERTembeddingused by

LaBSEembeddingused by

LASERembeddingused by

Universal Sentence Encoderembeddingused by

bge-multilingual-gemma2embeddingused by

bge-reranker-v2-m3rerankerused by

E5embeddingused by

Mistral Embedembeddingused by

Amazon Titan Text Embeddings V2embeddingused by

Cohere Embedembeddingused by

Embed English v3.0embeddingused by

Embed Multilingual v3.0embeddingused by

Embed v4.0embeddingused by

Embed v5.0 Proembeddingused by

Cohere Rerankrerankerused by

Rerank v3.5rerankerused by

Rerank v4.0 Fastrerankerused by

Rerank v4.0 Prorerankerused by

Gemini Embedding 2embeddingused by

OpenAI text embeddingsembeddingused by

text-embedding-3-largeembeddingused by

text-embedding-3-smallembeddingused by

text-embedding-ada-002embeddingused by

all-MiniLM-L12-v2embeddingused by

all-MiniLM-L6-v2embeddingused by

all-mpnet-base-v2embeddingused by

MS MARCO cross-encoders (MiniLM)rerankerused by

ms-marco-MiniLM-L12-v2rerankerused by

ms-marco-MiniLM-L6-v2rerankerused by

rerank-2.5rerankerused by

voyage-3.5embeddingused by

voyage-3-largeembeddingused by

voyage-4embeddingused by

voyage-code-3embeddingused by

Voyage embeddingsembeddingused by

Voyage rerankersrerankerused by

Papers

Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksPatrick Lewis et al., NeurIPS 2020 (per arXiv comment), 2020 founding
REALM: Retrieval-Augmented Language Model Pre-TrainingKelvin Guu et al., arXiv preprint, 2020 founding
Self-RAG: Learning to Retrieve, Generate, and Critique through Self-ReflectionAkari Asai et al., arXiv preprint, 2023 state of the art
Atlas: Few-shot Learning with Retrieval Augmented Language ModelsGautier Izacard et al., arXiv preprint, 2022 state of the art
Retrieval-Augmented Generation for Large Language Models: A SurveyYunfan Gao et al., arXiv preprint, 2023 survey

Where this comes from

The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

NVIDIA Nemotron nvidia.com https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/
Retrieval-Augmented Generation for Large Language Models: A Survey arxiv.org https://arxiv.org/abs/2312.10997

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Retrieval-Augmented Generation (RAG)", https://aurion.semai.id/topics/retrieval-augmented-generation/, records of October 2026.

Retrieval-Augmented Generation (RAG)

Tier Grades

Tier A: Frontier AI labs

Tier B: Indonesian AI labs

Tier C: Institution units such as ITB

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

Anchors

Who and What

Links

Models 130

Papers

Sources