Case studies / RAG and search

• ZZZenml14 Nov 2026NewRAG and search✓ Build details 2/5

  Bell: Building Modular and Scalable RAG Systems with Hybrid Batch/Incremental Processing

  Problem Documentation changed constantly, making it hard to keep the RAG system fast and up‑to‑date.

  Solution They built a modular hybrid RAG pipeline that runs batch jobs for static docs and incremental jobs for new docs, using Apache Beam, Cloud Composer and GCP services.

  • Hybrid search
  • Chunking
  • Embeddings

  DSuggested by Darshan Chougule

  Read on zenml.ioDetailsSave
• Elastic11 Nov 2026NewRAG and search

  Fine-grained authorization for Elasticsearch RAG with SpiceDB

  Problem Teams needed secure, fine‑grained access control for each document chunk in their Elasticsearch RAG chatbot.

  Solution They integrated SpiceDB to check each retrieved chunk against a relationship graph before passing it to the LLM, keeping authorization separate from the vector search.

  • Graph RAG
  • RAG
  • Vector search

  Read on elastic.coDetailsSave
• Hugging Face11 Nov 2026NewRAG and search

  Token Pooling-Aware Training for Multi-Vector Embedding Models

  Problem Storing millions of document vectors for multi‑vector search uses too much space.

  Solution They train the encoder to keep important information when similar token vectors are merged, letting the system compress vectors 4× while preserving retrieval quality.

  97.8% quality retention at 4× compression

  • RAG
  • Embeddings
  • Vector search

  Research behind this, on AI Papers DailyReducing Representational Interference in Retrieval Embeddings with Trainable Prefix Tokens

  Read on huggingface.coDetailsSave
• Apple5 Nov 2026RAG and search

  Faster Retrieval for Better, More Efficient RAG

  Problem Searching millions of document vectors was slow and costly, hurting response time for user questions.

  Solution Apple trained tiny neural networks offline to learn how typical queries map to the best document vectors, routing queries quickly to the right cluster for faster retrieval.

  • Routing
  • RAG
  • Hybrid search

  Read on machinelearning.apple.comDetailsSave
• JetBrains5 Nov 2026NewRAG and search

  Building a RAG Pipeline for Semantic Code Search: A Developer Diary and Field Notes

  Problem Developers needed instant, accurate code search across huge, constantly changing codebases.

  Solution JetBrains built Air Context, a production RAG system that chunks, vectors, stores millions of code vectors in an ANN index, supports incremental updates and branch‑level filtering for millisecond retrieval.

  • Vector search
  • RAG
  • Chunking

  Read on blog.jetbrains.comDetailsSave
• Qdrant23 Oct 2026RAG and search

  Scalable PDF Retrieval with Cohere Embed 5 and Qdrant Asymmetric Quantization

  Problem Searching PDF pages with images and complex layouts was slow and missed visual information.

  Solution They used Cohere Embed 5 to turn each PDF page image into a single dense vector and stored it in Qdrant with asymmetric quantization for fast, scalable retrieval.

  • Embeddings
  • Vector search
  • Reranking

  Research behind this, on AI Papers DailyContextual Embeddings Beyond the Gold Chunk

  Read on qdrant.techDetailsSave
• Elastic22 Oct 2026RAG and search

  56 scanned pages to a cited answer in 5 minutes: RAG pipeline for the documents search cannot read

  Problem Scanned PDFs and complex layouts were unreadable by traditional search, limiting access to important knowledge.

  Solution Elastic created a RAG pipeline that OCRs scanned pages into structured markdown, chunks and embeds the text, uses hybrid keyword‑semantic retrieval, and generates answers with page citations via Claude Haiku.

  • Hybrid search
  • RAG
  • Chunking

  Read on elastic.coDetailsSave
• Qdrant19 Oct 2026RAG and search

  UltraBERT Fits 32K Tokens in One Vector. RIP Chunking?

  Problem Long documents either lose context when split into small chunks or get cut off by embedding limits.

  Solution They created a 32K‑token UltraBERT model that embeds whole documents into one vector, compared it to traditional chunking in Qdrant, and added hybrid search and 1‑bit quantisation to reduce size.

  • Chunking
  • Vector search
  • Quantisation

  Research behind this, on AI Papers DailyContextual Embeddings Beyond the Gold Chunk

  Read on qdrant.techDetailsSave
• Booking.com19 Oct 2026RAG and search

  From Static Help Content to a RAG-Based Onboarding Agent for Partners

  Problem Partners get stuck in the long, confusing registration flow and need instant help without a human.

  Solution Built a chatbot inside the sign‑up flow that pulls help‑center articles, ranks them, and generates answers grounded in that content.

  • RAG
  • Chunking
  • Reranking

  Read on booking.aiDetailsSave
• Weaviate14 Oct 2026RAG and search

  PathSeer: Adaptive Filtered Vector Search

  Problem Filtered vector searches were slow and unreliable when filters changed frequently.

  Solution We added the PathSeer algorithm to the HNSW index, letting the system choose the best order for distance and filter checks, which speeds up filtered searches.

  Significant query‑time reductions

  • Hybrid search
  • Reranking
  • Vector search

  Read on weaviate.ioDetailsSave
• Qdrant13 Oct 2026RAG and search

  Perplexity's New Late-Interaction Models Meet Qdrant's 21x Vector Compression

  Problem Storing millions of document image vectors required huge storage and memory.

  Solution Qdrant applied 1‑bit TurboQuant compression to reduce vector size 21×, keeping 96.5% retrieval quality and using RAM for compressed vectors.

  96.5% quality retained

  • Quantisation
  • Vector search
  • RAG

  Read on qdrant.techDetailsSave
• Airbnb13 Oct 2026RAG and search

  CASTLE: training natural language search before the first real query

  Problem Airbnb needed natural‑language search but had no real user queries or relevance labels to train on.

  Solution They created CASTLE, a system that synthesises realistic search queries and relevance labels from booking data, then trains a retrieval model on this joint synthetic data.

  • RAG
  • Hybrid search
  • Query rewriting

  Read on medium.comDetailsSave
• Zalando11 Oct 2026RAG and search

  Automating Data Ingestion with a RAG Flywheel

  Problem Operations staff had to manually match many different supplier data schemas to a single internal format.

  Solution Created a RAG flywheel that stores past mappings, uses vector embeddings for semantic search, and an LLM to suggest new column mappings, with output constraints and pgvector for consistency.

  • RAG
  • Embeddings
  • Vector search

  Read on engineering.zalando.comDetailsSave
• Salesforce16 Sep 2026RAG and search

  How Data 360 Builds Trusted Context: The Enduring Layer for Enterprise AI

  Problem Need relevant, authorized enterprise data for each AI request without sending all data in the prompt.

  Solution Created Data 360 Trusted Context, a layer that selects the smallest needed data slice, checks permissions, and retrieves it on‑demand for LLM calls.

  • RAG
  • Query rewriting
  • Hybrid search

  Read on engineering.salesforce.comDetailsSave
• Pinterest11 Sep 2026RAG and search

  Evolving Pinterest’s Embedding Retrieval Platform

  Problem Embedding search was becoming too expensive and slow as the number of items grew into tens of billions.

  Solution They upgraded their internal search platform to use quantized integer vectors, hybrid queries and more flexible indexing, cutting memory use while keeping latency low.

  • Quantisation
  • Vector search
  • Hybrid search

  Read on medium.comDetailsSave
• TTThumbtack10 Sep 2026RAG and search

  Three principles for building a vector platform at Thumbtack

  Problem Teams found it hard to create and use vector search without extra infrastructure and high cost.

  Solution Thumbtack built a reusable vector platform that lets teams feed any embedding model, stores vectors in a database, and offers a simple API for similarity queries.

  • Vector search
  • RAG
  • Embeddings

  Read on medium.comDetailsSave
• Salesforce9 Sep 2026RAG and search

  Enterprise AI Accuracy: Building a More Trustworthy RAG Application

  Problem The RAG system often returned incorrect or garbled information, confusing users and risking trust.

  Solution They added validation steps, stricter retrieval filters, and post‑generation checks to ensure retrieved data matches the query and is formatted correctly before answering.

  • RAG
  • Guardrails
  • Reranking

  Read on engineering.salesforce.comDetailsSave
• Weaviate1 Sep 2026RAG and search

  How to extract meaning from charts and tables in PDFs

  Problem Users couldn’t find data hidden in charts or tables inside PDFs, so answers were wrong or missing.

  Solution They built a RAG system that embeds whole PDF pages as images, letting a late‑interaction model see charts and tables directly without extra OCR or chunking steps.

  • RAG
  • Vector search
  • Chunking

  Read on weaviate.ioDetailsSave
• Hugging Face21 Aug 2026RAG and search

  How Hugging Face Inference Endpoints, Jobs, and Buckets Power Search on Papers with Code

  Problem Researchers struggled to find relevant papers quickly, especially with vague queries, typos, or missing keywords.

  Solution Created a hybrid search system that mixes fast keyword lookup with semantic vector embeddings, combines results with reciprocal rank fusion, and uses rerankers for better relevance.

  Search covers over 110,000 papers

  • Hybrid search
  • RAG
  • Reranking

  Read on huggingface.coDetailsSave
• Hugging Face18 Aug 2026RAG and search

  Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers

  Problem Single-vector embeddings lose detail, hurting search accuracy for specific queries.

  Solution They added a multi‑vector model that stores a vector for each token, matching queries token‑by‑token with a MaxSim operation, improving retrieval quality.

  • RAG
  • Embeddings
  • Vector search

  Read on huggingface.coDetailsSave
• Weaviate13 Aug 2026RAG and search

  Building Foundry Part 2: Where creative workflows break

  Problem Retrieval stopped working as projects grew and metadata quality slipped.

  Solution They built Foundry on Weaviate, using semantic embeddings and vector search to keep files findable despite chaotic folder and tagging practices.

  • RAG
  • Embeddings
  • Vector search

  Read on weaviate.ioDetailsSave
• Hugging Face12 Aug 2026RAG and search

  Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

  Problem Users needed a simple, cheap way to turn satellite images into numeric vectors for search and analysis.

  Solution OlmoEarth Studio now lets users run a model on chosen images and download the resulting embedding vectors as lightweight GeoTIFF files, with options for resolution, time range, and fine‑tuning.

  • Embeddings
  • Vector search
  • RAG

  Research behind this, on AI Papers DailyReducing Representational Interference in Retrieval Embeddings with Trainable Prefix Tokens

  Read on huggingface.coDetailsSave
• Weaviate11 Aug 2026RAG and search

  Scaling Test-Time Compute in Search Mode

  Problem Complex queries needed better search accuracy, but users also cared about speed and cost.

  Solution Added an effort setting that lets the query agent spend more time breaking down queries and reranking results, with three tiers from medium to ultrahigh.

  nDCG@10 57.5 vs 13.0

  • Reranking
  • Query rewriting
  • Hybrid search

  Read on weaviate.ioDetailsSave
• Booking.com11 Aug 2026RAG and search

  How we selected the next vector database at Booking.com

  Problem Teams needed a fast, scalable vector store that could handle many use cases and high traffic.

  Solution They built a shared vector‑search infrastructure, evaluated different databases, and chose a new system that supports hybrid search, large capacity, and high request rates.

  • Vector search
  • Hybrid search

  Read on booking.aiDetailsSave