Buyer's Guide: Enterprise Search & Agentic RAG Platforms

Executive Summary

Enterprise RAG is retrieval-augmented generation run over a company’s own content. A search step finds the passages that answer a question, and a language model writes the answer from them, with citations. An arXiv abstract describes retrieval-augmented generation (RAG) models as models which combine pre-trained parametric and non-parametric memory for language generation. In an enterprise, the non-parametric memory is an index of documents, tickets, messages and records, and the work is keeping that index current and matched to who may read what.

Agentic RAG adds planning to the search step. Instead of running one query, a model splits a question into several, runs them, and merges the results, or an agent calls search as one tool among several. Microsoft describes agentic retrieval in Azure AI Search as a multi-query pipeline designed for complex questions posed by users or agents in chat and copilot apps. Microsoft states that agentic retrieval adds latency compared to a single-query pipeline, but handles query complexity that a single query cannot. The same trade applies to every vendor: more steps mean more time and more model calls per answer.

No platform here is the right answer for every buyer, and this guide does not rank them. The products differ on four things you can test: whether they keep a copy of your content or search it in place, how they apply the permissions of each source system, whether they ship a finished assistant or the parts to build one, and what the price list counts. The vendor section groups eight products by those differences.

Why Permissions Decide Whether RAG Is Safe to Deploy

A RAG system is only as careful as its retrieval step. If the index returns a passage the user should not see, the model summarizes it. The OWASP GenAI Security Project warns that inadequate or misaligned access controls can lead to unauthorized access to embeddings containing sensitive information, and that the model could then retrieve and disclose personal data, proprietary information, or other sensitive content. OWASP recommends fine-grained access controls and permission-aware vector and embedding stores, with strict logical and access partitioning of datasets in the vector database.

Permission-aware retrieval also inherits every mistake in the source systems. A folder shared with the whole company is searchable by the whole company, and an assistant makes it easy to find. Microsoft advises using the permission models available in Microsoft 365 services, such as SharePoint, to help ensure the right users or groups have the right access to the right content. Clean up oversharing in the largest repositories before a broad rollout, whichever product you choose.

Retrieved content is also an attack path. OWASP states that indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files. OWASP describes an attack in which a document modified in a repository used by a RAG application is returned for a query and its instructions alter the output of the LLM. The exposure grows when the model can act as well as answer, because a planted instruction can then start an action.

Agents reach content through a protocol as well as through connectors. The Model Context Protocol project describes MCP as an open-source standard for connecting AI applications to external systems. Several products in this guide use MCP in one direction or the other: to pull content from other systems, or to let outside agents query their own index. Ask which direction a vendor means, and whether permissions are checked on the MCP path the same way as in its own interface.

Grounding can be measured. A grounded answer says only what the retrieved passages support, and a checker model can score that. Vectara publishes HHEM-2.1-Open, an open source model in its HHEM series for detecting hallucinations in LLMs. Whatever you buy, keep a set of questions with known answers from your own content, and score each release against it.

Assistant, Retrieval Layer or Parts: What You Are Buying

Enterprise RAG is sold at three depths. A finished assistant gives employees a search box and a chat window over connected systems. A retrieval layer returns ranked, permission-checked passages to an application or agent you build. A set of parts (a search engine, a vector index, an embedding model) leaves the pipeline and the permission logic to your team.

The finished assistant is the quickest to roll out and the hardest to shape. The parts are the most flexible and put the security work on you. The retrieval layer sits between them: the vendor maintains connectors and permissions, and you own the experience. Large organizations often run more than one, such as an employee assistant from their productivity suite and a separate retrieval service behind a customer-facing agent.

The other split is whether content is copied. Glean states that with indexed access, content and permissions are crawled earlier, stored in your Glean tenant and read at query time, and Glean applies the permission snapshot that was mirrored during crawling and incremental updates. Glean states that with live access, Glean calls the source system during the user request to fetch current records or fields. An index answers quickly and can rank results across every system, but its permissions are only as current as the last sync. A live call is current, but it is slower, bounded by the source’s API limits, and harder to rank against results from other systems.

How do you evaluate Enterprise Search & Agentic RAG Platforms?

Permission handling comes first. Ask how each connector reads access control lists, groups and sharing links from its source, how long a revoked permission takes to reach search results, and what happens to content whose permissions cannot be read. On a search engine you build on, this logic is yours. Elastic states that you can control access to data within a data stream or index by adding field and document level security permissions to a role. Mapping each source system’s groups to those roles, and keeping them in sync, is then work for your team.

Retrieval quality sets answer quality. Keyword search finds exact terms, part numbers and names. Vector search finds passages with similar meaning in different words. Elastic states that hybrid search runs full-text search and vector search in one request. Elastic recommends implementing hybrid search with the reciprocal rank fusion (RRF) algorithm. Ask each vendor whether search is hybrid, whether results are reranked, and how you can measure relevance on your own queries.

Connectors set reach and freshness. Count the systems you need rather than the size of the vendor’s catalog, and test each one for content types, permissions and the delay between a change at the source and a change in results. For systems without a connector, ask for an API to push content in, and who maintains a custom connector when the source changes.

Citations make answers checkable. Every generated answer should link to the passages it used, and the user should be able to open them with their own access. Ask what the system does when retrieval finds nothing relevant, and test it with questions your content cannot answer.

Agent support is the newest layer. Ask whether outside agents can call the platform as a tool, whether an agent acts with the permissions of the user who started it, and which actions require approval. Glean states that an agent should not automatically access everything the user can. Check which agent features are generally available. Google Cloud states that some capabilities, including delegating multi-step background tasks to Gemini, are currently in early access.

Vendors in Enterprise Search & Agentic RAG Platforms

Representative vendors include Amazon Q Business (Amazon Quick), Coveo, Elastic, Glean, Google Gemini Enterprise, Microsoft 365 Copilot, Sinequa by ChapsVision and Vectara. The list is alphabetical, is not a ranking, and is not for sale. How we choose vendors

Eight products are grouped by two questions: does the product ship a finished assistant or a retrieval layer for others to build on, and where does its index live?

A work assistant with its own index is the first approach. Glean states that it supports 275+ out-of-the-box connectors across enterprise apps and ecosystems. Glean states that teams can bring custom data and actions into Glean with custom connectors, the Indexing SDK, MCP servers, and OpenAPI-based actions.

Productivity suites and cloud providers sell assistants that ground on the content you already keep with them, and connect to other systems. Microsoft states that Copilot only surfaces organizational data to which individual users have at least view permissions. Microsoft states that synced connectors index data into Microsoft Graph, while federated connectors use a Model Context Protocol (MCP) model to fetch data in real time without indexing content into Microsoft 365. Microsoft states that the Microsoft 365 Copilot Retrieval API lets you ground your own generative AI solutions by returning relevant text chunks from the hybrid index that powers Microsoft 365 Copilot. Google Cloud describes Gemini Enterprise as an intranet search, AI assistant, and agentic platform. Google Cloud states that Gemini indexes selected sources while keeping data where it lives, and that access strictly follows the existing permissions of each user. Google Cloud lists over 100 supported integrations, including Google Workspace, Microsoft 365, Slack, Confluence, Jira, Salesforce and ServiceNow. Google Cloud states that you can connect data sources to Gemini Enterprise using a custom Model Context Protocol (MCP) server.

One product in this approach is closed to new buyers. Amazon Web Services states that Amazon Q Business is no longer available to new customers. Amazon Web Services describes Amazon Quick as the next evolution of Amazon Q Business. Amazon Web Services states that existing Q Business customers can continue using their current service or use their existing Q index with Quick. Amazon Web Services states that Amazon Q Business can understand and respect your existing identities, roles, and permissions. New buyers evaluate its successor; existing customers decide whether and when to move.

A search engine you build on is the third approach. Elastic describes Elasticsearch as an open source, distributed search and analytics engine that stores structured, unstructured, and vector data in real time. Elastic announced that it would add AGPL as another license option next to ELv2 and SSPL. Elastic states that Elastic Agent Builder lets you create custom agents that use your data through skills, tools, and prompts. Connectors, permission mapping and the answer interface are then decisions for your team.

A relevance platform sits between a finished assistant and a search engine. Coveo states that its content security option named Same users and groups as in your content system secures content by replicating the item permission system of the repository. Coveo states that a Passage Retrieval (CPR) model retrieves the passages from your indexed enterprise content that are most relevant to a query. Coveo states that the generated answer includes citations that reference the indexed content used to generate the answer. Coveo states that its Hosted Model Context Protocol (MCP) Server lets LLM applications and AI agent frameworks access its enterprise search and retrieval capabilities.

Search for complex and restricted estates is the fifth approach. Sinequa states that it is now part of the Argonos Operational AI Platform. ChapsVision states that its AI products, including Sinequa and Systran, are part of the Argonos platform. Sinequa states that it grounds AI agents, AI assistants, and intelligent search in trusted internal data while enforcing document-level security at enterprise scale. Argonos states that source-system permissions are inherited and kept in sync, so that changes in access at the source are reflected in Argonos. Argonos states that it can be deployed on-premises as well as in cloud environments.

A RAG platform behind an API is the sixth. Vectara describes itself as an API-first Agentic Platform for building production-grade AI agents and assistants that retrieve knowledge, reason over context, use tools, and deliver grounded answers. Vectara states that it supports SaaS, VPC, on-premises, and air-gapped deployments.

Eight products are named across six approaches, and three share the suite and cloud provider approach. The approaches were written before any product was selected, and a row in this table is not for sale. Other enterprise search engines, open-source stacks and assistants built into business applications are in the market, and any of them can make its own case in the Spotlight below.

How much should you budget for Enterprise Search & Agentic RAG Platforms?

Enterprise RAG is priced per user, per unit of indexed or stored content, per query, or by model and compute consumption, and many price lists combine two of these. Price the meter that grows with your usage rather than the entry price.

Index capacity can be its own line. Amazon Web Services states that, with Amazon Q Business, enterprise customers pay for user subscriptions and index capacity. Amazon Web Services states that once an index is created, you incur charges for it regardless of whether you have used any document capacity within that index. Vectara states that its Context Storage Utilization metric captures the size of the data stored within the Vectara platform for retrieval and agent execution. Corpus size, not head count, drives these charges, so index the sources people search rather than whole repositories.

Generated answers and passage retrieval can be extensions to a search license. Coveo states that Relevance Generative Answering (RGA) is a paid product extension. Coveo states that Passage Retrieval (CPR) is a paid product extension. Ask which features in a demo require an extension, and price them at your expected query volume.

Model choice changes the bill as well. Google Cloud states that included monthly seat credits apply strictly to first-party Gemini models, and that partner model usage is billed as pay-as-you-go usage. Agentic retrieval multiplies model calls per question, so estimate cost from the number of steps per answer as well as from the number of users.

A search engine you build on replaces a per-user fee with infrastructure, embedding and engineering costs. Price the people who will maintain connectors, permission mapping and relevance tuning, and the model calls that generate answers.

How long does implementation take for Enterprise Search & Agentic RAG Platforms?

Roll out by permission readiness, not by connector count. Start with content whose access is already clean, prove that answers respect it, then add sources and audiences.

What should you ask vendors about Enterprise Search & Agentic RAG Platforms?

Ask each vendor to answer in writing and to demonstrate on your own content, accounts and test questions.

Related Resources