AI Search Results Practice Exam

Question 1 AI Search Results

How does AI search handle ambiguous queries?

By using context clues, clarification questions, and multiple intent detection to provide relevant results or ask for clarification when needed. Multimodal AI search can understand and retrieve information across text, images, audio, and video formats, not just text-based queries. Retrievers must filter results to documents the current user can access; solutions include passing ACL as metadata filters, indexing with user/group claims, and post-filtering on access tokens. Metadata filters (e.g. tenant_id, date > 2024, language = "en") are pre-filters that restrict ANN search to a subset of vectors, critical for multi-tenant and time-bounded retrieval.

Question 2 AI Search Results

What is SigLIP and how does it improve on CLIP?

SigLIP replaces CLIP's softmax contrastive loss with a sigmoid loss, scaling better and achieving higher zero-shot ImageNet accuracy at smaller batch sizes; used in Gemini vision encoders. Perplexity's Sonar model uses a search agent that iteratively queries the web, reads pages, refines the query, and synthesizes a cited answer, exposed via the Sonar API. A Vectara benchmark tracking hallucination rate vs. summary length; GPT-4 class models run at ~3% hallucination, while earlier models exceeded 20%. ELSER is Elastic's out-of-domain learned sparse retriever that ships as a managed model in Elasticsearch, giving BM25-style sparse vectors with semantic understanding.

Question 3 AI Search Results

What is Cohere Rerank and what is its typical API signature?

co.rerank(query=..., documents=[...], top_n=k, model="rerank-english-v3.0"); it returns indices and relevance scores used to reorder first-stage hits before the LLM step. OpenAI, Anthropic, and Google now offer prompt caching that reuses a cached prefix (e.g. system prompt + retrieved docs) for 5-10 minutes, cutting cost up to 90% and latency up to 80%. BM25 is a probabilistic term-frequency ranking function; it remains a strong baseline and is often hybridized with neural retrievers because it excels at exact-match recall on rare terms. Techniques that allow search systems to handle new types of queries without extensive retraining, adapting quickly to new domains and languages.

Question 4 AI Search Results

What is query rewriting in RAG and why does it matter?

Query rewriting uses an LLM to transform a conversational or ambiguous query into a self-contained search query (e.g. resolving pronouns, expanding acronyms), dramatically improving recall on multi-turn dialogues. A common default is 512 tokens with 10-20% overlap (~50-100 tokens), with chunking tuned per content type; too-small chunks lose context, too-large chunks dilute relevance. Metadata filters (e.g. tenant_id, date > 2024, language = "en") are pre-filters that restrict ANN search to a subset of vectors, critical for multi-tenant and time-bounded retrieval. Massive Text Embedding Benchmark (MTEB) covers 56 datasets across retrieval, classification, clustering, and semantic textual similarity, and is the de facto leaderboard for general-purpose embeddings.

Question 5 AI Search Results

What role does Information Retrieval play in AI development?

Information Retrieval (IR) is a cornerstone of AI, enabling systems to efficiently find, access, and process relevant data from vast datasets. A Vectara benchmark tracking hallucination rate vs. summary length; GPT-4 class models run at ~3% hallucination, while earlier models exceeded 20%. Pinecone is a managed serverless vector database (built on proprietary infrastructure) popular for production RAG, offering metadata filtering, namespaces, and hybrid sparse+dense search. Streaming returns LLM tokens as they are generated (server-sent events), so the user sees the answer build up in real time even when the full answer takes 5-10 seconds to complete.

Question 6 AI Search Results

How does AI search market growth compare to traditional search?

BM25 is a probabilistic term-frequency ranking function; it remains a strong baseline and is often hybridized with neural retrievers because it excels at exact-match recall on rare terms. Generative AI transforms search by understanding specific inputs including Boolean terms, connectors, and filters, providing more contextually relevant results. SimpleQA is OpenAI's short-form factuality benchmark with single-answer questions, designed to measure hallucination rates of frontier models on factual queries. AI search has grown to capture 7.82% of the search market by mid-2025, showing significant but still emerging adoption compared to established engines.

Question 7 AI Search Results

How does dense retrieval differ from sparse retrieval in modern AI search?

Sparse retrieval (e.g. BM25) matches exact lexical tokens, while dense retrieval encodes queries and documents into vector embeddings and matches by cosine similarity, capturing semantic meaning. Fan-out = the agent splits a question into sub-questions, retrieves in parallel, and aggregates results — the pattern underlying Deep Research and multi-hop agentic RAG. Long-context LLMs fit everything in the prompt but are expensive, slow, and still miss mid-document facts; RAG retrieves only relevant spans and is cheaper, faster, and more auditable per query. 1) Missing content (not retrieved), 2) Missed top-ranked docs, 3) Not in context (lost in middle), 4) Not extracted (LLM ignores the answer even when present).

Question 8 AI Search Results

What is Google's Speech-to-Retrieval engine?

Projections suggest AI search will grow significantly through 2030, potentially competing equally with or surpassing traditional search engines. A 2025 Google innovation that directly interprets voice search and retrieves information without intermediate text conversion, enabling more natural voice-based search interactions. AI search understands context, intent, and semantic meaning rather than just matching keywords, providing more relevant and personalized results. Schema.org is a vocabulary of structured data (FAQ, Product, Article) embedded in HTML; AI search engines parse it for entity disambiguation, prices, dates, and citation-ready facts, boosting GEO inclusion.

Question 9 AI Search Results

What is product quantization (PQ) in vector search?

Metadata filters (e.g. tenant_id, date > 2024, language = "en") are pre-filters that restrict ANN search to a subset of vectors, critical for multi-tenant and time-bounded retrieval. A website can include invisible text ("Ignore prior instructions and...") that an AI search agent ingests as instructions, exfiltrating user data or changing behavior — a top OWASP LLM risk in 2025. PQ splits each embedding into sub-vectors and replaces each with the nearest centroid id from a codebook, compressing vectors by ~32x with small recall loss for billion-scale retrieval. Leo is Brave's private, no-log AI answer assistant inside the browser; Brave's premium Leo subscription (~$15/mo) and its ad-free index differentiate it from Google-dependent competitors.

Question 10 AI Search Results

What is You.com and how does it differ from Perplexity?

Weaviate is an open-source vector DB with built-in vectorization modules (OpenAI, Cohere, HuggingFace) and GraphQL/REST APIs, supporting hybrid search and generative RAG out of the box. Hierarchical Navigable Small World (HNSW) is a graph-based ANN index that builds multi-layer proximity graphs; it offers logarithmic-time search, high recall, and incremental inserts. You.com offers a multi-modal, multi-agent answer engine with customizable apps (YouAgents) and modes (Smart, Genius, Research), emphasizing customization over Perplexity's minimal UI. Site owners can block GPTBot (OpenAI), ClaudeBot (Anthropic), Google-Extended, CCBot, and PerplexityBot via standard robots.txt; User-agent: GPTBot\nDisallow: / blocks training and retrieval.

Question 11 AI Search Results

How does AI search impact user privacy concerns?

WebArena (2023) is a realistic, self-hostable web agent benchmark built on clones of GitHub, Reddit, Maps, and shopping sites; top agents still complete <30% of tasks in 2025. Semantic caching stores previous (query, embedding, answer) tuples and returns a cached answer if a new query is within a similarity threshold of an old one — useful for high-volume, repetitive queries. AI search raises concerns about data collection for personalization, query storage, and profiling, leading to debates about privacy vs. personalization benefits. Perplexity Pages (2024) lets users turn a search session into a shareable, citable article with images, sections, and inline sources, blurring the line between search and publishing.

Question 12 AI Search Results

What is Milvus / Zilliz?

DPR (Karpukhin et al., 2020) uses two BERT encoders (question + passage) trained with contrastive loss on QA pairs; the first dense retriever to outperform BM25 on open-domain QA. Milvus is an open-source, distributed vector DB built by Zilliz with multiple index types (HNSW, IVF, DiskANN) targeting billion-scale deployments, the OSS counterpart to Pinecone at scale. AI search UIs put the generated answer at the top with citations, followed by sources, related questions, and traditional links — the inverse of classic "10 blue links" SERPs. Answer provenance tracks the exact document, version, and span that contributed each claim, so audits can reproduce the answer — required in finance, pharma, and legal AI deployments.

Question 13 AI Search Results

What is the typical latency budget for a production RAG query?

The choice of cosine, dot product, or Euclidean affects ranking; cosine is standard for normalized embeddings, dot product for unnormalized, Euclidean for absolute-distance clustering. LLMs give more weight to context at the beginning and end of the prompt and ignore the middle, so reranking should place the most relevant chunks at the top and bottom of the context window. Query rewriting uses an LLM to transform a conversational or ambiguous query into a self-contained search query (e.g. resolving pronouns, expanding acronyms), dramatically improving recall on multi-turn dialogues. ~1-3 seconds end-to-end: ~50-150ms retrieval, ~100-300ms reranking, ~500-2000ms LLM streaming first token. Anything >3s drops user engagement sharply.

Question 14 AI Search Results

What content licensing deals has OpenAI signed with publishers?

OpenAI, Microsoft, and Google signed content licensing agreements with AP, Axel Springer, News Corp, the Financial Times, The Atlantic, Vox, Le Monde, and Reddit (~$60M/yr) for both training data and live retrieval rights. Long-context LLMs fit everything in the prompt but are expensive, slow, and still miss mid-document facts; RAG retrieves only relevant spans and is cheaper, faster, and more auditable per query. Perplexity's Sonar model uses a search agent that iteratively queries the web, reads pages, refines the query, and synthesizes a cited answer, exposed via the Sonar API. AI systems that can search and retrieve information across multiple languages, understanding queries in one language and finding relevant content in others.

Question 15 AI Search Results

What is a knowledge graph in AI search?

llms.txt is a proposed markdown index of a site's high-quality LLM-readable content, placed at /llms.txt, so AI crawlers can ingest a concise site summary instead of crawling thousands of HTML pages. CRAG adds a lightweight retrieval evaluator that grades retrieved docs as Correct / Ambiguous / Incorrect and triggers web search or fallback for low-confidence cases. Typical Deep Research tasks cost $0.50 - $5 in API (or 1-25 Pro credits) due to 100+ web fetches and long reasoning traces — 10-100x a simple RAG query. A knowledge graph (KG) stores entities and relations (e.g. (Tesla, founded_by, Elon_Musk)); retrieval over a KG supports symbolic, multi-hop reasoning that pure vector search cannot.

Question 16 AI Search Results

What is step-back prompting in RAG?

Through External Information Retrieval capabilities that access current data from the web, APIs, and live sources beyond training data. A website can include invisible text ("Ignore prior instructions and...") that an AI search agent ingests as instructions, exfiltrating user data or changing behavior — a top OWASP LLM risk in 2025. It generates a more abstract or general query alongside the original, retrieves context for both, and feeds both into the LLM — improving answers on factual, multi-hop, or scientific questions. ColBERTv2 adds residual compression and denormalized embeddings, cutting index size by ~10x and improving retrieval quality over the original ColBERT.

Question 17 AI Search Results

What is the difference between ReAct and Chain-of-Thought?

Techniques that allow search systems to handle new types of queries without extensive retraining, adapting quickly to new domains and languages. Schema.org is a vocabulary of structured data (FAQ, Product, Article) embedded in HTML; AI search engines parse it for entity disambiguation, prices, dates, and citation-ready facts, boosting GEO inclusion. RAG combines retrieval of relevant documents with generative AI to produce accurate, context-aware answers grounded in retrieved information. CoT reasons in natural language only; ReAct interleaves external tool calls (search, calculator, code) with reasoning, grounding answers in live evidence.

Question 18 AI Search Results

What are the key characteristics of advanced AI search solutions in 2025?

They deliver faster, smarter, and more accurate results using machine learning, natural language processing, and context understanding. It generates a more abstract or general query alongside the original, retrieves context for both, and feeds both into the LLM — improving answers on factual, multi-hop, or scientific questions. Generative AI transforms search by understanding specific inputs including Boolean terms, connectors, and filters, providing more contextually relevant results. The forecast extends from 2025 to 2030, projecting when ChatGPT and AI-powered search will surpass traditional search engines like Google.

Question 19 AI Search Results

How do businesses prepare for AI search dominance?

Chinese AI search products are integrating Generative AI with conventional internet search, creating hybrid search solutions tailored to the Chinese market. AI search understands context, intent, and semantic meaning rather than just matching keywords, providing more relevant and personalized results. Query rewriting uses an LLM to transform a conversational or ambiguous query into a self-contained search query (e.g. resolving pronouns, expanding acronyms), dramatically improving recall on multi-turn dialogues. By optimizing content for semantic relevance, structured data, and question-answering rather than traditional keyword SEO strategies.

Question 20 AI Search Results

What is the role of small language models (SLMs) in AI search?

A capability that allows AI agents to access new information unseen during training, enabling continuous learning and up-to-date knowledge retrieval. Site owners can block GPTBot (OpenAI), ClaudeBot (Anthropic), Google-Extended, CCBot, and PerplexityBot via standard robots.txt; User-agent: GPTBot\nDisallow: / blocks training and retrieval. Announced on Pixel 8 and Samsung Galaxy S24 in 2024, Circle to Search lets users circle any on-screen object to trigger a visual AI search combining image and text context. SLMs (Phi-4, Gemma 2 9B, Llama 3.1 8B) increasingly handle query rewriting, routing, and reranking at the edge, leaving only the final generation to a frontier model — cutting cost 10-50x.

Question 21 AI Search Results

Compare HNSW, IVF-PQ, and ScaNN in one sentence each.

Query rewrite → hybrid retrieve (BM25 + dense) → rerank → top-k context → LLM with citations → answer, with optional HyDE, query expansion, and post-generation grounding checks. HNSW = graph-based, high recall, more RAM; IVF-PQ = inverted-file + product quantization, low memory, lower recall; ScaNN = Google's anisotropic vector quantization, top-tier on TPU/CPU benchmarks. Toolformer (Meta, 2023) is a model that self-supervises API calls by inserting them into text and training only when the call actually improves perplexity, enabling web search, calculator, and QA tools. LongLLMLingua uses a small model to remove non-essential tokens from retrieved documents (e.g. adjectives, restatements), achieving 4-20x compression with minimal quality loss.

Question 22 AI Search Results

What are multimodal search capabilities in AI?

DiskANN is a Microsoft graph-based ANN index designed for SSD-resident vectors at billion scale, with Vamana graph; used in Azure AI Search and Bing's vector backend. SPLADE learns sparse but expanded representations using BERT's MLM head, predicting which vocabulary terms should be activated per document — combining neural semantics with lexical sparsity. Multimodal AI search can understand and retrieve information across text, images, audio, and video formats, not just text-based queries. Deep Research is a multi-step research agent that browses 50-200+ sources, runs code, and produces a long-form report with citations in 5-30 minutes, launched across major AI search products in early 2025.

Question 23 AI Search Results

What is HNSW and why is it the most popular vector index?

CoT reasons in natural language only; ReAct interleaves external tool calls (search, calculator, code) with reasoning, grounding answers in live evidence. SigLIP replaces CLIP's softmax contrastive loss with a sigmoid loss, scaling better and achieving higher zero-shot ImageNet accuracy at smaller batch sizes; used in Gemini vision encoders. Hierarchical Navigable Small World (HNSW) is a graph-based ANN index that builds multi-layer proximity graphs; it offers logarithmic-time search, high recall, and incremental inserts. By learning from user behavior, context, and preferences to deliver results tailored to individual needs and search patterns.

Question 24 AI Search Results

What is "permissions-aware" retrieval?

Retrievers must filter results to documents the current user can access; solutions include passing ACL as metadata filters, indexing with user/group claims, and post-filtering on access tokens. MRL trains embeddings so that the first d dimensions form a useful embedding (e.g. 64, 128, 256, 512), allowing one model to serve multiple storage/recall budgets without retraining. Multiple analyses (Authoritas, 2024) estimate AI Overviews and ChatGPT search could cut publisher organic traffic 20-60% for informational queries, prompting affiliate- and direct-traffic pivots. Glean is an AI search and assistant for the enterprise that indexes Slack, Google Drive, Confluence, GitHub, Salesforce, etc., respects per-document ACLs, and grounds answers in source links.

Question 25 AI Search Results

What is "answer provenance" in enterprise AI search?

Fan-out = the agent splits a question into sub-questions, retrieves in parallel, and aggregates results — the pattern underlying Deep Research and multi-hop agentic RAG. Many enterprise answers live in SQL, knowledge graphs, or APIs, not documents. Modern AI search combines text-to-SQL, KG-RAG, and tool-calling agents to retrieve structured facts alongside unstructured passages. Answer provenance tracks the exact document, version, and span that contributed each claim, so audits can reproduce the answer — required in finance, pharma, and legal AI deployments. Hybrid search combines BM25 + dense vector retrieval (and sometimes cross-encoder reranking) to get lexical precision and semantic recall, with reciprocal rank fusion (RRF) merging the two ranked lists.

Question 26 AI Search Results

What is Toolformer?

Toolformer (Meta, 2023) is a model that self-supervises API calls by inserting them into text and training only when the call actually improves perplexity, enabling web search, calculator, and QA tools. Query rewrite → hybrid retrieve (BM25 + dense) → rerank → top-k context → LLM with citations → answer, with optional HyDE, query expansion, and post-generation grounding checks. They deliver faster, smarter, and more accurate results using machine learning, natural language processing, and context understanding. OpenAI, Anthropic, and Google now offer prompt caching that reuses a cached prefix (e.g. system prompt + retrieved docs) for 5-10 minutes, cutting cost up to 90% and latency up to 80%.

Question 27 AI Search Results

Name three state-of-the-art open embedding models as of 2025.

Their 2023/2024 study coined Generative Engine Optimization (GEO), showing techniques like citing sources, adding statistics, and using quotations can boost source visibility in LLM answers by up to 40%. AI systems that can search and retrieve information across multiple languages, understanding queries in one language and finding relevant content in others. BGE-M3, E5-Mistral-7B, and Nomic-Embed-Text-v1.5 are leading open embeddings, with MTEB benchmark scores above 65 for English and strong multilingual performance. Semantic caching stores previous (query, embedding, answer) tuples and returns a cached answer if a new query is within a similarity threshold of an old one — useful for high-volume, repetitive queries.

Question 28 AI Search Results

What is the market share of ChatGPT as a "search engine" in 2025?

A 2025 Google innovation that directly interprets voice search and retrieves information without intermediate text conversion, enabling more natural voice-based search interactions. ChatGPT receives an estimated ~5-6% of global search referrals and is the most-used "AI search destination", per Similarweb and SparkToro analyses in 2025. Semantic caching stores previous (query, embedding, answer) tuples and returns a cached answer if a new query is within a similarity threshold of an old one — useful for high-volume, repetitive queries. OpenAI's WebGPT (2021) showed that a GPT-3 model fine-tuned to browse the web and cite sources could produce answers preferred over human-written Reddit responses, foreshadowing modern AI search.

Question 29 AI Search Results

What is Maximal Marginal Relevance (MMR) in retrieval?

MMR reorders retrieved documents to maximize relevance while minimizing redundancy, scoring each candidate by λ * sim(q, d) - (1-λ) * max sim(d, d') for d' already selected. Training crawls (e.g. GPTBot, CCBot) ingest content for model training; retrieval crawls (e.g. OAI-SearchBot, ChatGPT-User) fetch pages live to ground answers and obey per-page fetch policies. Multiple analyses (Authoritas, 2024) estimate AI Overviews and ChatGPT search could cut publisher organic traffic 20-60% for informational queries, prompting affiliate- and direct-traffic pivots. Embeddings convert text into numerical vector representations that capture semantic meaning, enabling AI to find conceptually similar content beyond keyword matches.

Question 30 AI Search Results

What is Perplexity Pages?

Multimodal AI search can understand and retrieve information across text, images, audio, and video formats, not just text-based queries. Microsoft Copilot combines GPT-4-class models with Bing's index to deliver cited, conversational answers, and is integrated across Edge, Windows, and Microsoft 365. LanceDB is an embedded, serverless vector DB built on the Lance columnar format, optimized for ML workloads (training data, embeddings) with zero ops. Perplexity Pages (2024) lets users turn a search session into a shareable, citable article with images, sections, and inline sources, blurring the line between search and publishing.

Question 31 AI Search Results

What technologies enable modern AI search engines?

Milvus is an open-source, distributed vector DB built by Zilliz with multiple index types (HNSW, IVF, DiskANN) targeting billion-scale deployments, the OSS counterpart to Pinecone at scale. Neural networks, machine learning, natural language processing (NLP), vector embeddings, and transformer models power modern AI search capabilities. Vespa is an open-source engine from Yahoo for large-scale hybrid search + ranking + ML inference, used by companies needing tens of thousands of QPS with sub-100ms latency. LongLLMLingua uses a small model to remove non-essential tokens from retrieved documents (e.g. adjectives, restatements), achieving 4-20x compression with minimal quality loss.

Question 32 AI Search Results

How do vector databases support AI search?

ColPali embeds document page images directly with a vision-language model (PaliGemma) and uses ColBERT-style late interaction, beating OCR-then-text pipelines on visually rich PDFs. Vector databases store and efficiently retrieve embeddings, enabling fast semantic search across millions of documents using similarity matching. Advanced AI search systems provide direct citations and links to original sources, allowing users to verify information and explore further. ANN trades a small loss in recall for huge speed gains over exact kNN by using index structures like HNSW, IVF-PQ, or ScaNN, enabling sub-linear vector lookup over millions of embeddings.

Question 33 AI Search Results

What is the difference between enterprise AI search and consumer AI search?

Pinecone is a managed serverless vector database (built on proprietary infrastructure) popular for production RAG, offering metadata filtering, namespaces, and hybrid sparse+dense search. Enterprise AI search requires access control (ACLs), audit logs, connectors to SharePoint/Confluence/Slack, hybrid indexes, and provenance; consumer AI search optimizes for freshness, coverage, and cost. Microsoft Copilot combines GPT-4-class models with Bing's index to deliver cited, conversational answers, and is integrated across Edge, Windows, and Microsoft 365. Training crawls (e.g. GPTBot, CCBot) ingest content for model training; retrieval crawls (e.g. OAI-SearchBot, ChatGPT-User) fetch pages live to ground answers and obey per-page fetch policies.

Question 34 AI Search Results

What is ELSER (Elastic Learned Sparse EncodeR)?

ELSER is Elastic's out-of-domain learned sparse retriever that ships as a managed model in Elasticsearch, giving BM25-style sparse vectors with semantic understanding. Hierarchical Navigable Small World (HNSW) is a graph-based ANN index that builds multi-layer proximity graphs; it offers logarithmic-time search, high recall, and incremental inserts. Fan-out = the agent splits a question into sub-questions, retrieves in parallel, and aggregates results — the pattern underlying Deep Research and multi-hop agentic RAG. HyDE prompts an LLM to generate a hypothetical answer first, embeds that, and retrieves documents similar to the hypothetical — useful for zero-shot domains where the query is short.

Question 35 AI Search Results

What is agentic RAG?

OCR errors propagate to embeddings and BM25 indexes, hurting recall on tables, equations, handwriting, and low-resolution scans; vision-native retrievers like ColPali skip OCR entirely. GraphRAG builds a knowledge graph from documents (entities + relations), then retrieves subgraphs relevant to the query, often using community detection to answer global questions. Users expect <2 seconds for AI answers; production systems use streaming (start typing at ~300ms), reranker cascading, and prompt caching to meet this bar at scale. Agentic RAG gives the LLM tools and a loop to decide when and what to retrieve, when to re-query, when to call APIs, and when to answer — replacing the rigid retrieve-then-generate pipeline.

Question 36 AI Search Results

What is semantic caching in AI search?

Metadata filters (e.g. tenant_id, date > 2024, language = "en") are pre-filters that restrict ANN search to a subset of vectors, critical for multi-tenant and time-bounded retrieval. Agentic RAG gives the LLM tools and a loop to decide when and what to retrieve, when to re-query, when to call APIs, and when to answer — replacing the rigid retrieve-then-generate pipeline. Semantic caching stores previous (query, embedding, answer) tuples and returns a cached answer if a new query is within a similarity threshold of an old one — useful for high-volume, repetitive queries. BM25 is a probabilistic term-frequency ranking function; it remains a strong baseline and is often hybridized with neural retrievers because it excels at exact-match recall on rare terms.

Question 37 AI Search Results

What are the main challenges facing AI search adoption?

pgvector is a PostgreSQL extension that adds vector types and ANN search (HNSW, IVF) to Postgres, popular for teams that want a single OLTP+vector store without a new system. MCP (Anthropic, 2024) is a standardized JSON-RPC protocol for connecting LLMs to tools and data sources (search, files, DBs), enabling plug-and-play AI search integrations across clients and vendors. A knowledge graph (KG) stores entities and relations (e.g. (Tesla, founded_by, Elon_Musk)); retrieval over a KG supports symbolic, multi-hop reasoning that pure vector search cannot. Challenges include accuracy concerns, computational costs, integration with existing systems, and user trust in AI-generated results.

Question 38 AI Search Results

What is the typical p95 latency for hybrid search at 10M vectors?

Many enterprise answers live in SQL, knowledge graphs, or APIs, not documents. Modern AI search combines text-to-SQL, KG-RAG, and tool-calling agents to retrieve structured facts alongside unstructured passages. With HNSW on modern hardware, p95 hybrid search (BM25 + dense) over 10M vectors runs in ~50-150ms, well within the RAG latency budget. Embedding too many chunks per query wastes tokens and dilutes attention; better to retrieve, then compress or re-rank down to the 5-15 most relevant chunks before the LLM. CLIP (Contrastive Language-Image Pretraining) trains a text and image encoder in a shared space using 400M (image, caption) pairs, enabling zero-shot image search and text-to-image retrieval.

Question 39 AI Search Results

What is the difference between traditional search and AI-powered search?

Traditional search relies on keyword matching and PageRank, while AI search uses semantic understanding and machine learning to interpret queries and results. 1) Missing content (not retrieved), 2) Missed top-ranked docs, 3) Not in context (lost in middle), 4) Not extracted (LLM ignores the answer even when present). Glean is an AI search and assistant for the enterprise that indexes Slack, Google Drive, Confluence, GitHub, Salesforce, etc., respects per-document ACLs, and grounds answers in source links. Through technologies like Google's Speech-to-Retrieval engine that directly converts spoken language to search intent without intermediate text processing.

Question 40 AI Search Results

What is Google's AI Overviews (formerly SGE)?

llms.txt is a proposed markdown index of a site's high-quality LLM-readable content, placed at /llms.txt, so AI crawlers can ingest a concise site summary instead of crawling thousands of HTML pages. SPLADE learns sparse but expanded representations using BERT's MLM head, predicting which vocabulary terms should be activated per document — combining neural semantics with lexical sparsity. AI Overviews are generative summaries that appear above traditional blue-link results in Google Search, launched broadly in May 2024 and powered by a customized Gemini model. Fan-out = the agent splits a question into sub-questions, retrieves in parallel, and aggregates results — the pattern underlying Deep Research and multi-hop agentic RAG.

Question 41 AI Search Results

What is the typical RAG pipeline in 2025?

Neural networks, machine learning, natural language processing (NLP), vector embeddings, and transformer models power modern AI search capabilities. Glean is an AI search and assistant for the enterprise that indexes Slack, Google Drive, Confluence, GitHub, Salesforce, etc., respects per-document ACLs, and grounds answers in source links. LongLLMLingua uses a small model to remove non-essential tokens from retrieved documents (e.g. adjectives, restatements), achieving 4-20x compression with minimal quality loss. Query rewrite → hybrid retrieve (BM25 + dense) → rerank → top-k context → LLM with citations → answer, with optional HyDE, query expansion, and post-generation grounding checks.

Question 42 AI Search Results

What is the typical architecture of a 2025 enterprise RAG application?

Challenges include accuracy concerns, computational costs, integration with existing systems, and user trust in AI-generated results. BM25 is a probabilistic term-frequency ranking function; it remains a strong baseline and is often hybridized with neural retrievers because it excels at exact-match recall on rare terms. Connectors (M365, Slack, Drive) → chunker → embedding + BM25 index → ACL-aware retriever → reranker → grounded LLM with citations → audit log + provenance, with separate evaluation and feedback loops. DPR (Karpukhin et al., 2020) uses two BERT encoders (question + passage) trained with contrastive loss on QA pairs; the first dense retriever to outperform BM25 on open-domain QA.

Question 43 AI Search Results

Why did the New York Times sue OpenAI and Microsoft?

Lexical search matches exact words and phrases, while semantic search understands meaning and concepts, finding relevant results even without exact keyword matches. MRL trains embeddings so that the first d dimensions form a useful embedding (e.g. 64, 128, 256, 512), allowing one model to serve multiple storage/recall budgets without retraining. The NYT filed suit in December 2023 alleging that ChatGPT and Copilot reproduce its articles verbatim and that OpenAI used its content to build a substitute product, seeking statutory and actual damages. A website can include invisible text ("Ignore prior instructions and...") that an AI search agent ingests as instructions, exfiltrating user data or changing behavior — a top OWASP LLM risk in 2025.

Question 44 AI Search Results

What is token streaming and why does it matter?

Fan-out = the agent splits a question into sub-questions, retrieves in parallel, and aggregates results — the pattern underlying Deep Research and multi-hop agentic RAG. A capability that allows AI agents to access new information unseen during training, enabling continuous learning and up-to-date knowledge retrieval. Streaming returns LLM tokens as they are generated (server-sent events), so the user sees the answer build up in real time even when the full answer takes 5-10 seconds to complete. Through External Information Retrieval capabilities that access current data from the web, APIs, and live sources beyond training data.

Question 45 AI Search Results

How does generative AI impact traditional search outcomes?

Perplexity's Sonar model uses a search agent that iteratively queries the web, reads pages, refines the query, and synthesizes a cited answer, exposed via the Sonar API. Sparse retrieval (e.g. BM25) matches exact lexical tokens, while dense retrieval encodes queries and documents into vector embeddings and matches by cosine similarity, capturing semantic meaning. AI search UIs put the generated answer at the top with citations, followed by sources, related questions, and traditional links — the inverse of classic "10 blue links" SERPs. Generative AI transforms search by understanding specific inputs including Boolean terms, connectors, and filters, providing more contextually relevant results.

Question 46 AI Search Results

What is the role of context in AI search results?

BM25 is a probabilistic term-frequency ranking function; it remains a strong baseline and is often hybridized with neural retrievers because it excels at exact-match recall on rare terms. ~1-3 seconds end-to-end: ~50-150ms retrieval, ~100-300ms reranking, ~500-2000ms LLM streaming first token. Anything >3s drops user engagement sharply. AI search uses conversation history, user preferences, and search context to provide more relevant and personalized results over time. Projections suggest AI search will grow significantly through 2030, potentially competing equally with or surpassing traditional search engines.

Question 47 AI Search Results

What is the main risk of over-embedding in RAG?

Projections suggest AI search will grow significantly through 2030, potentially competing equally with or surpassing traditional search engines. Embedding too many chunks per query wastes tokens and dilutes attention; better to retrieve, then compress or re-rank down to the 5-15 most relevant chunks before the LLM. Query rewriting uses an LLM to transform a conversational or ambiguous query into a self-contained search query (e.g. resolving pronouns, expanding acronyms), dramatically improving recall on multi-turn dialogues. Lexical search matches exact words and phrases, while semantic search understands meaning and concepts, finding relevant results even without exact keyword matches.

Question 48 AI Search Results

Define hybrid search and explain why it is the dominant production pattern.

The OpenSearch neural_search plugin adds dense and sparse neural retrievers (via external models served in OpenSearch Dashboards) with hybrid scoring and rerankers integrated into the DSL. Hybrid search combines BM25 + dense vector retrieval (and sometimes cross-encoder reranking) to get lexical precision and semantic recall, with reciprocal rank fusion (RRF) merging the two ranked lists. BM25 is a probabilistic term-frequency ranking function; it remains a strong baseline and is often hybridized with neural retrievers because it excels at exact-match recall on rare terms. DPR (Karpukhin et al., 2020) uses two BERT encoders (question + passage) trained with contrastive loss on QA pairs; the first dense retriever to outperform BM25 on open-domain QA.

Question 49 AI Search Results

What is "fan-out" in agentic AI search?

Fan-out = the agent splits a question into sub-questions, retrieves in parallel, and aggregates results — the pattern underlying Deep Research and multi-hop agentic RAG. Facebook AI Similarity Search (FAISS) is Meta's open-source library for efficient similarity search and clustering of dense vectors, with GPU support and many index types — the toolkit beneath most vector DBs. AI search traffic accounts for 7.82% of the combined search market as of June 2025. Answer provenance tracks the exact document, version, and span that contributed each claim, so audits can reproduce the answer — required in finance, pharma, and legal AI deployments.

Question 50 AI Search Results

How do AI search engines combat misinformation?

By using source verification, citation linking, confidence scoring, and retrieval from authoritative sources to ensure accuracy. LLMs give more weight to context at the beginning and end of the prompt and ignore the middle, so reranking should place the most relevant chunks at the top and bottom of the context window. Connectors (M365, Slack, Drive) → chunker → embedding + BM25 index → ACL-aware retriever → reranker → grounded LLM with citations → audit log + provenance, with separate evaluation and feedback loops. MMR reorders retrieved documents to maximize relevance while minimizing redundancy, scoring each candidate by λ * sim(q, d) - (1-λ) * max sim(d, d') for d' already selected.