Map of Content: Agentic Search & Embeddings

Entry point for the semantic knowledge graph covering the Agentic Search and Embeddings sections of the Awesome Search collection. Navigate by concept, person, or article.

---

Core Paradigms

• Agentic Search — agents that plan, execute, verify, and refine retrieval
• RAG — retrieval-augmented generation pipeline
• Semantic Search — meaning-based retrieval
• Hybrid Search — sparse + dense combination

---

Embedding Architectures

Architecture	Key Property	Note
Bi-Encoder	Independent encoding, fast	Two-tower, dual encoder
Cross-Encoder	Joint encoding, accurate	Reranker
ColBERT	Per-token MaxSim	Late interaction
Late Interaction	Token-level alignment	ColBERT’s mechanism

---

Vector Types

Dense

• Embeddings — what embeddings are; dense vs sparse; training; economics

• Dense Embeddings — all-dimension active vectors; bi-encoder models; ANN indexing

• Dense Vector Retrieval — ANN indexes (HNSW, IVF, Flat)

• Matryoshka Embeddings — flexible-dimension, nested quality

• Task-Aware Embeddings — instruction-guided representations

• Multimodal Embeddings — cross-modal (image + text)

• Vector Similarity Metrics — cosine similarity, dot product, Euclidean (L2) distance

Sparse

• Sparse Embeddings — vocabulary-space vectors; BM25 vs learned sparse (SPLADE/ELSER)
• Sparse Vector Retrieval — inverted index compatible
• SPLADE — BERT MLM → vocabulary-space sparse vectors
• ELSER — Elastic’s production SPLADE

Constructed Query Vectors

• Hypothetical Document Embeddings — LLM generates hypothetical answer → embed
• Wormhole Vectors — bridge sparse/dense/behavioral spaces
• Bag-of-Documents Model — query as distribution over documents

Quantization

• Vector Quantization — compressing embeddings: scalar (SQ8), binary (BBQ), product (PQ), K-Quants/I-Quants
• BBQ — Elasticsearch’s Better Binary Quantization + OSQ; 32× compression, 10–40× speed
• GGUF — quantized LLM weight format for local deployment (rerankers, query expansion)

---

Retrieval Strategies

• Retrieval Pipeline — multi-stage: retrieve → rerank
• Asymmetric Semantic Search — short query → long document
• Vector Filtering — metadata predicates + ANN search
• Embedding Fine-tuning — domain adaptation

ANN Index Structures

• HNSW — graph-based; best recall/speed; dominant in Elasticsearch, Qdrant, Weaviate
• IVF — cluster-based; lower memory; IVF-PQ for billion-scale

Quantization Methods

• Scalar Quantization — int8/int4 per coordinate; 4–8× compression; universal
• Binary Quantization — 1-bit per coordinate; 32× compression; needs rescoring

Dimensionality Reduction

• Topic: Dimensionality Reduction vs Quantization — two multipliers on bytes = dims × bits/dim; how to stack them
• Topic: PCA vs t-SNE for Retrieval — why visualization DR and retrieval DR are different tasks
• PCA — orthogonal linear projection; the retrieval-safe default
• Matryoshka Embeddings — DR moved into model training; truncation free at query time

---

Text Preparation

• Text Chunking — fixed, recursive, semantic, contextual methods

---

Key People

Person	Affiliation	Key Contributions
Daniel Tunkelang	QueryUnderstanding.com	Agentic Search, Bag-of-Documents, Pipeline Distillation
Omar Khattab	Stanford	ColBERT creator
Matei Zaharia	Stanford/Databricks	ColBERT co-creator
Jo Kristian Bergum	Hornet (co-founder); ex-Vespa	ColBERT 32x compression; agentic query workload
Han Xiao	Jina AI	jina-colbert-v1-en (8192 tokens)
Trey Grainger	—	Wormhole Vectors, “AI-Powered Search”
Stéphane Clinchant	NAVER LABS	SPLADE co-inventor
Thibault Formal	NAVER LABS	SPLADE co-inventor
James Briggs	Pinecone	SPLADE explainer, Vector Filtering
Shaw Talebi	—	Fine-tuning text & multimodal embeddings
Asif Makhani	Infino AI	Co-authored Agentic Search with Tunkelang
Dima Kan	Aiven	Wormhole Vectors implementation
Lester Solbakken	hornet.dev	MAD framework, defensive retrieval
Skip Everling	hornet.dev	Scaling dimensions of keyword search; agentic query workload
Thomas Veasey	Elastic	BBQ/OSQ quantization benchmarks
Ivan Pleshkov	Qdrant	TurboQuant + RaBitQ implementation in Qdrant 1.18
Piotr Mazurek	tensoreconomics.com	Embeddings economics, FLOPS/dollar analysis
Quynh Nguyen	Elastic	Multilingual embedding hybrid search

---

Articles by Topic

Agentic Search

• Agentic Search as an Agile Engineering Process

• Agentic search models

• Agents turn simple keyword search into compelling search experiences

• Superintelligent Retrieval Agent SIRA — LLM-enriched BM25; corpus enrichment + query expansion; beats agentic RAG on BEIR

• Mutually Assured Distraction — Lester Solbakken; MAD dynamic; abstention as retrieval control signal

• You Say Search I Say Recs - Spotify Agentic Query Understanding — LLM router; +115% similar artists, +91% new music releases

• Incremental AI Adoption for E-commerce Search — 4-level progression: traditional → conversational AI

• This Is What Agentic Retrieval Looks Like — Jo Kristian Bergum; GPT-5 BrowseComp-Plus analysis; median 24 searches/session; phrase quotes in 98%; agent queries at 99th percentile of human length

• The Scaling Dimensions of Keyword Search — Skip Everling (Hornet); serving-side cost of long agent queries; Block-Max WAND degrades past ~7 terms

• Agentic Search Models with OpenSearch and Elasticsearch — Bonsai; SID-1 purpose-built agentic search/rerank model dropped into OpenSearch

Frontier topic: Frontier of Search 2026 — two converging fronts: search systems for agents + Purpose-Built Agentic Search Models. Underlying shift: Agentic Query Workload.

ColBERT / Late Interaction

• Announcing the Vespa ColBERT Embedder
• What is ColBERT and Late Interaction and Why They Matter in Search?
• Bi-encoder vs Cross-encoder: When to Use Which One

Matryoshka Embeddings

• Matryoshka Embeddings - Faster OpenAI Vector Search
• Introduction to Matryoshka Embedding Models 1
• Matryoshka Representation Learning - A Guide to Faster Semantic Search 1

Text Chunking

• Chunking Strategies for LLM Applications
• Evaluating the Ideal Chunk Size for a RAG System using LlamaIndex 1
• How to Chunk Text Data - A Comparative Analysis 1

Context-Aware / Task-Aware Embeddings

• Improve your RAG applications by moving to Task-aware Embeddings
• How Context-Aware Embeddings Are Transforming Enterprise Search

SPLADE & Sparse Retrieval

• Hybrid Search SPLADE Sparse Encoder
• SPLADE for Sparse Vector Search Explained
• SPLADE - sparse bi-encoder BERT model for first-stage ranking 1
• Elastic Learned Sparse Encoder (ELSER) Retrieval Performance

Constructed Query Vectors

• Hypothetical Document Embeddings (HyDE)

• Distilling Retrieval Pipelines to a Single Embedding Model

• Beyond Hybrid Search - Traversing Vector Spaces with Wormhole Vectors

• Wormhole Vectors Beyond Hybrid Search in OpenSearch — Dima Kan; production SKG traversal at Aiven

Fine-tuning

• Fine-Tuning Text Embeddings For Domain-Specific Search 1

• Fine-tuning Multimodal Embedding Models 1

• Fine Tune LLM on a Custom Dataset with QLoRA

• Fine-Tuning LLMs From Zero to Hero with Python and Ollama

• Fine-Tuning Qwen3 Embeddings for Product Category Classification — LoRA on 615M; LSPC dataset; 0.836 macro-F1

• Qwen3 Embedding Series — 0.6B/4B/8B; #1 MTEB multilingual; MRL support

Vector Infrastructure

• Nearest Neighbor Indexes for Similarity Search 1

• The Missing WHERE Clause in Vector Search 1

• How to Choose the Best Model for Semantic Search

• Symmetric vs Asymmetric Semantic Search

• Choosing the best model for semantic search

• From Zero to Semantic Search Embedding Model

• Why Are Embeddings So Cheap — Piotr Mazurek; compute-bound; ~$0.01/1M tokens at scale

• Multilingual Embedding Model Hybrid Search Reranking — Quynh Nguyen; E5 multilingual + RRF + Cohere reranker

Quantization

• Elasticsearch BBQ Optimized Scalar Quantization vs TurboQuant — Thomas Veasey; OSQ vs TurboQuant; 10–40× CPU speedup via integer SIMD
• TurboQuant in Qdrant — Ivan Pleshkov & Jonas Schulz; Qdrant 1.18; rotation-based quantization with anisotropy compensation; beats BQ by 9–24 pp recall at same storage
• GGUF Quantization - A Technical Deep Dive — Michael Hannecke; K-Quants, I-Quants, imatrix, deployment guide

Case Studies

• Migrating to Elasticsearch with dense vector for Carousell Spotlight 1
• Innovating Search Experience with Amazon OpenSearch and Amazon Bedrock 1

---

Concept Relationship Map

Agentic Search
  └── uses → Retrieval Pipeline
               ├── Stage 1: Bi-Encoder / Sparse Vector Retrieval / Hybrid Search
               └── Stage 2: Cross-Encoder / ColBERT (Late Interaction)

RAG
  └── uses → Dense Vector Retrieval
               ├── input: Text Chunking → Bi-Encoder embeddings
               └── query: Asymmetric Semantic Search / HyDE / Task-Aware Embeddings

Hybrid Search
  ├── sparse leg: SPLADE / ELSER / BM25
  └── dense leg: Bi-Encoder
       └── advanced: Wormhole Vectors / Bag-of-Documents Model

Embedding Quality
  ├── Matryoshka Embeddings (flexible dimensions)
  ├── Embedding Fine-tuning (domain adaptation)
  ├── Multimodal Embeddings (image + text)
  └── Vector Filtering (metadata + ANN)

New: Direct Corpus Interaction

• Beyond Semantic Similarity - Rethinking Retrieval for Agentic Search via Direct Corpus Interaction — TIGER-Lab; DCI (grep/bash) outperforms all traditional retrievers; +30.7% multi-hop QA, +21.5% IR ranking
• Metadata - The 3rd Kind of Retrieval — Doug Turnbull; attribute-based retrieval as a 3rd paradigm; LLMs make attribute extraction easy

New: ColBERT Training

• ColBERT-Zero - To Pre-train Or Not To Pre-train ColBERT Models — Antoine Chaffin et al. (LightOn AI); multi-vector pre-training beats KD-only; SOTA BEIR <150M (55.43 nDCG@10)
• ColBERT Comes to Apache Solr — Nicolò Rinaldi (Sease); ColBERT reranking implementation for Apache Solr