Enterprise Search
Enterprise search in 2026: open-source and commercial options compared

Shekhar Kadyan
Co-founder

Enterprise search is software that lets employees search every company app, file and message from one place, and shows each person only the results they are allowed to see. Enterprise AI search adds a language model on top, so people can ask a question in plain words and get a written answer that cites the documents it came from.
Disclosure: PipesHub publishes this article, and PipesHub is one of the products compared below. We have tried to give every option a fair hearing and to credit each rival's real strengths. Every claim about another company links to a dated source. Competitor facts are as of November 2026.
Key takeaways
Permissions are the hard part. Any tool can index documents. The tools worth buying check, at the moment someone searches, that this person may see each result.
Good answers need more than one kind of search. Modern systems combine keyword search with meaning-based search, reorder the results by relevance, and cite the exact passages they used.
"Open source" covers very different things. Some projects are finished workplace search products, and others are engines you build on. Licences differ too, so check which features sit in a paid "enterprise" folder.
Where it runs matters as much as what it does. You can use the vendor's cloud, have the vendor manage it inside your cloud, host it yourself, or run it air-gapped, which means with no internet connection at all.
Agents are now a second audience. The Model Context Protocol (MCP) is an open standard that lets AI assistants use outside tools and data. Through it, the search layer that answers employees can also serve assistants such as Claude Code or Cursor, with the same permission checks.
The category is now formally tracked. The analyst firm Gartner published its first Magic Quadrant, its vendor comparison report, for Enterprise AI Search in October 2026. It assessed 11 vendors (Elastic via Business Wire, 7 Oct 2026).
How modern enterprise AI search works
A modern system has eight parts. When you evaluate a product, ask how it handles each one.

The path of one question. Content and access rules are synced into the index beforehand.
1. Connectors and sync
A connector is a small program that reads data from one app, such as Slack, Google Drive, Jira or SharePoint. It copies that data into the search system's index, which is the store the system searches.
Connectors differ in what they bring across. Some bring only the text, while others also bring attachments and comments. The most important thing they bring is the record of who has access.
Connectors also differ in how fresh their data is. Some sync on a schedule, and others pick up changes within minutes.
2. Parsing and indexing
First, each document is converted to text. This includes tables and scanned pages, which are read with OCR (optical character recognition).
The text is then split into passages called chunks. Each chunk is stored twice. One copy goes into a keyword index. The other is stored as an embedding, a list of numbers that captures what the passage means, so that similar passages can be found later.
3. Permission enforcement at query time
Every company app has its own ACLs, or access control lists. These are the rules that say who may open what. A good enterprise search tool copies those rules into its index and filters every search by them. That way an intern never sees a board deck, even if it matches their question.
The details matter. If permissions are copied only once a day, a person removed from a folder this morning may still see its files this afternoon. Ask each vendor how quickly permission changes take effect, and whether the check happens when the query runs.
4. Hybrid retrieval
Keyword search is excellent at exact matches such as error codes, product names and ticket numbers. It often uses a ranking formula called BM25. Vector search compares embeddings, so it finds passages with the same meaning even when the words differ.
Hybrid retrieval runs both and merges the results. This means neither kind of question is missed.
5. Knowledge graph
A knowledge graph stores the relationships between things. For example, it records that this pull request fixes that Jira ticket, and that the ticket was discussed in this Slack thread by these people. Retrieval can follow those links to find context that a plain text match would miss.
6. Reranking
The first retrieval step is built for speed and returns a broad set of candidates. A reranker is a second, slower model that reads the question and each candidate together. It reorders the candidates by relevance, so the best passages reach the language model.
7. Cited answers
The language model writes an answer from the top passages only, and it links each statement to its source. Citations let people check the answer. They are also the main defence against mistakes that sound confident, which are often called hallucinations.
8. Agents and MCP
The Model Context Protocol (MCP) is an open standard that lets AI assistants call outside tools and data. When an enterprise search system offers an MCP server, assistants such as Claude, Cursor or Gemini CLI can search company knowledge through it.
The key question is whose permissions the agent uses. It should act as a specific person, not as a service account that can see everything.
Deployment options: SaaS, your cloud, self-hosted, air-gapped
SaaS (software as a service) means the vendor runs the product for you in its own cloud. The table below compares it with the other three ways to run enterprise search.
Model | Who runs it | Where data lives | Typical fit |
|---|---|---|---|
SaaS | The vendor | The vendor's cloud | Teams that want the fastest start and the least operations work |
Vendor-managed in your cloud | The vendor, inside your AWS or GCP account | Your cloud account | Teams that need their data to stay in their own account without running the system themselves |
Self-hosted | You, on your own servers or cloud | Your infrastructure | Teams that want control over data, models and cost, and are ready to handle upgrades |
Air-gapped | You, on a network with no internet | Your isolated network | Defence, government and regulated research, where the language models must run locally |
An air-gapped setup needs every component to run locally, including the language model and the embedding model. Not every product that says "self-hosted" can do this, so ask for the air-gapped install guide.
Evaluation checklist you can reuse
Copy this list into your request for proposal or pilot plan.
Connectors: Does it cover every system that holds your knowledge? For each one, does it bring permissions, attachments and comments, and how fresh is the data?
Permission model: Are source permissions enforced on every query? How long does a revoked permission take to apply? Can you test it with two real users?
Retrieval quality: Does it use hybrid retrieval and reranking? Run 50 of your own real questions and score the answers.
Citations: Does each answer link to the exact passage, not just the file?
Model choice: Can you bring your own language and embedding models, including local ones?
Deployment: Which of SaaS, your cloud, self-hosted and air-gapped are supported and documented?
Licence: Which licence covers the code you would run? Which features sit in a paid or "enterprise" edition? These are often single sign-on, audit logs or permission sync.
Identity: Which tier includes single sign-on (through the SAML or OIDC standards), automatic user provisioning and audit logs?
Agents and APIs: Is there an MCP server, an API and SDKs (software development kits)? Do agents run with a person's own permissions?
Operations: What hardware does it need? What is the upgrade path? Does it support Kubernetes, backup and monitoring?
Evidence: Are there public benchmarks, an evaluation you can reproduce, and customer references you can call?
Total cost: What will the licence or seat price, infrastructure, model usage and staff time add up to over three years?
The 2026 landscape: ten options compared
These facts are as of November 2026, and the sources are listed at the end. "Self-host" means you can run the product on infrastructure you control.
Option | Licence | Self-host? | Real strengths | Best for |
|---|---|---|---|---|
Glean | Proprietary | Partly. Its "Customer Hosted" option runs in your own AWS or GCP account and is managed by Glean. | Scale and polish. It passed $300M in annual recurring revenue in May 2026, and it offers a wide connector catalogue, an assistant and agents. | Large companies wanting a fully managed product |
Microsoft 365 Copilot / Microsoft Search | Proprietary | No | It is built into Microsoft 365 and uses its existing permissions. Copilot is a paid add-on, listed at $30 per user per month, paid yearly (Microsoft, checked 20 Nov 2026). | Organisations whose knowledge lives mostly in Microsoft 365 |
Atlassian Rovo | Proprietary | No, it runs in Atlassian Cloud. A Data Center connector lets Cloud Rovo read on-premises Jira and Confluence. | It is included in paid Atlassian Cloud plans and has deep Jira and Confluence context. | Teams that live in Jira and Confluence Cloud |
Coveo | Proprietary | None published. Coveo describes itself as a "single SaaS platform", and its platform page mentions no self-hosted option (checked 20 Nov 2026). | Relevance tuning. It is a Leader in Gartner's 2025 Magic Quadrant for Search and Product Discovery, and it offers a hosted MCP server. | Customer-service and website search, with workplace search alongside |
Sinequa (ChapsVision) | Proprietary | Yes. It runs on-premises, in a private cloud or as a hybrid of the two (vendor-reported). | A long track record in regulated industries | Large regulated enterprises wanting a commercial on-premises product |
Guru | Proprietary | Not published. Guru's pricing page says nothing about self-hosting (checked 20 Nov 2026). | Verified, curated knowledge and governance. It has about 2,322 G2 reviews and delivers its knowledge through MCP. | Teams that value curated, verified answers over raw search |
Elastic or OpenSearch build | Elasticsearch is offered under AGPL, SSPL or Elastic License 2.0. OpenSearch uses Apache 2.0. | Yes | Both are proven search engines at very large scale. Elastic was named a Leader in Gartner's first Enterprise AI Search Magic Quadrant. It lists 33 ready-made connectors, many with permission sync (Elastic docs, checked 20 Nov 2026). | Teams with search engineers who want to build their own answer layer and user interface on a proven engine |
Onyx | MIT outside its "ee" folders, and the Onyx Enterprise License inside them | Yes, including air-gapped (vendor-reported) | It has the largest community among open-source workplace search apps, with 33k+ GitHub stars. It runs EnterpriseRAG-Bench, and its cloud plan starts at $20 per user per month. | Teams wanting a popular open-source chat and search app |
RAGFlow | Apache 2.0 | Yes | About 94k GitHub stars and strong document parsing | Developers building their own retrieval apps rather than deploying a finished workplace search product |
PipesHub (publisher) | Apache 2.0 | Yes, including a documented air-gapped install | It checks permissions at query time, combines a knowledge graph with hybrid retrieval, gives cited answers, includes an MCP server and offers 50+ connectors. | Teams that need self-hosted, permission-aware search and want the same context for their agents |
Two notes on Onyx: automatic permission syncing is an Enterprise Edition feature, and Onyx announced on 21 Oct 2026 that it is joining Zoom.
We have not read Gartner's paywalled report. The placements mentioned here come from the vendors' own announcements.
Where PipesHub fits, and where it doesn't
Where it fits. PipesHub is an open-source project under the Apache 2.0 licence. You can run it with one install command on Docker Compose, a tool that starts a set of containers together. You can also run it on Kubernetes, a system for running containers at scale, using the published Helm chart. A managed option, PipesHub Cloud, is available too.
PipesHub connects to 50+ systems, including Slack, Google Drive, Jira, Confluence, Microsoft 365, GitHub and Notion. It resolves access when the query runs, against each source system's own permissions. Answers come back with block-level citations, which point to the exact block of the original document that supports each statement.
For retrieval, PipesHub combines a knowledge graph with hybrid retrieval. The graph is stored in Neo4j or ArangoDB. The hybrid retrieval runs in the Qdrant vector database and uses both dense vectors, which capture meaning, and sparse vectors, which capture exact terms.
You can bring any language model, including local models through Ollama, and the docs include an air-gapped installation guide. The remote MCP server lets Claude Code, Cursor, Gemini CLI, Claude.ai and LibreChat search company knowledge as a specific person. There are also SDKs for Python, TypeScript and Go.
Connectors sync on a schedule, and real-time sync is available for some connectors in the Enterprise edition. Single sign-on and audit logs are part of the Enterprise edition (pricing).
On evidence, PipesHub ranked #1 on the public EnterpriseRAG-Bench leaderboard (Hugging Face) as of 14 Nov 2026, with 93.24 overall (95.0% answer correctness) among 43 entries (leaderboard). Onyx created and runs this benchmark and leaves its own product off the leaderboard. It is a synthetic test set, so treat it as one signal and run your own questions too.
Who should choose something else.
A company whose knowledge lives almost entirely in one suite. If nearly everything is in Microsoft 365 or Atlassian Cloud, the bundled assistant may be good enough and cheaper to adopt.
A team that wants to build search from parts. Elasticsearch, OpenSearch and RAGFlow are engines and toolkits made for that.
A large company that wants a fully managed, proprietary product. Glean is built for that buyer.
FAQ
What is the difference between enterprise search and enterprise AI search?
Enterprise search returns a ranked list of documents from across company apps, filtered by each person's permissions. Enterprise AI search keeps that permission-aware retrieval but adds a language model that reads the top results and writes a direct answer with citations to the sources.
Is there an open-source alternative to Glean?
Yes. Onyx and PipesHub are open-source workplace search and assistant platforms that you can self-host, and RAGFlow, Elasticsearch and OpenSearch are open engines you can build on. Compare their licences carefully, because some keep features such as permission syncing or single sign-on in a paid edition.
Can enterprise AI search run fully self-hosted or air-gapped?
Yes, if every component runs locally. Self-hosted products such as PipesHub, Onyx and Sinequa can run on your own infrastructure. An air-gapped install also needs a local language model and a local embedding model, because the network cannot reach hosted AI providers.
How does enterprise search stop people from seeing documents they shouldn't?
Connectors copy each source app's access rules alongside the content, and every search is filtered by those rules for the person asking. The best systems check access when the query runs, so a revoked permission takes effect quickly rather than after the next full sync.
What is hybrid search, and why does it matter?
Hybrid search combines keyword search, which finds exact terms such as error codes, with vector search, which finds passages with similar meaning. Merging both and then reranking the results usually finds the right passage more often than either method alone.
How do AI agents use enterprise search through MCP?
The Model Context Protocol lets an AI assistant call an enterprise search system as a tool. The assistant sends a question, receives permission-filtered passages with citations, and uses them to answer or act. The agent should sign in as a real person so it sees only what that person can see.
Did Gartner publish a Magic Quadrant for enterprise AI search?
Yes. Gartner published its first Magic Quadrant for Enterprise AI Search in October 2026, assessing 11 vendors. Elastic announced it was named a Leader, and Squirro and IntraFind announced Niche Player placements.
Related reading
Sources
All accessed 14–20 November 2026.
Elastic, "Elastic Recognized as a Leader in the Gartner Magic Quadrant for Enterprise AI Search", Business Wire, 7 Oct 2026: https://www.businesswire.com/news/home/20261007553151/en/Elastic-Recognized-as-a-Leader-in-the-Gartner-Magic-Quadrant-for-Enterprise-AI-Search
Squirro, Gartner Magic Quadrant announcement: https://squirro.com/news-and-events/squirro-gartner-magic-quadrant-enterprise-ai-search-2026 ; IntraFind: https://intrafind.com/en/news/intrafind-positioned-in-2026-gartner-magic-quadrant-for-enterprise-ai-search
Glean, $300M ARR, Business Wire, 28 May 2026: https://www.businesswire.com/news/home/20260528505530/en/Glean-Surpasses-$300M-ARR-Unrivaled-Enterprise-Context-Fuels-AI-Adoption
Glean deployment models (Glean Hosted, Customer Hosted): https://docs.glean.com/get-started/prepare/about-deployment
Microsoft 365 Copilot for enterprise, pricing: https://www.microsoft.com/en-us/microsoft-365-copilot/enterprise
Atlassian Rovo licensing: https://www.atlassian.com/licensing/rovo ; Rovo pricing: https://atlassian.com/software/rovo/pricing
Coveo, Leader in 2025 Gartner Magic Quadrant for Search and Product Discovery: https://ir.coveo.com/en/news-events/press-releases/detail/440/coveo-named-a-leader-in-the-2025-gartner-magic ; Coveo hosted MCP server: https://ir.coveo.com/en/news-events/press-releases/detail/474/coveo-announces-hosted-mcp-server-to-expand-enterprise-ai ; pricing page: https://www.coveo.com/en/pricing
Sinequa by ChapsVision: https://www.chapsvision.com/platform/sinequa/ ; acquisition, Business Wire, 18 Nov 2024: https://www.businesswire.com/news/home/20241118042171/de
Guru pricing and package: https://www.getguru.com/pricing ; Guru G2 reviews: https://www.g2.com/products/Guru/reviews
Elastic licensing FAQ: https://www.elastic.co/pricing/faq/licensing ; "Elasticsearch is open source again": https://www.elastic.co/blog/elasticsearch-is-open-source-again ; OpenSearch repository (Apache 2.0): https://github.com/opensearch-project/OpenSearch
Onyx licence: https://github.com/onyx-dot-app/onyx/blob/main/LICENSE ; pricing: https://onyx.app/pricing ; llms.txt (Enterprise Edition features, air-gapped deployments): https://onyx.app/llms.txt ; "Onyx is joining Zoom", 21 Oct 2026: https://onyx.app/blog/onyx-joins-zoom ; GitHub: https://github.com/onyx-dot-app/onyx
RAGFlow repository (Apache 2.0): https://github.com/infiniflow/ragflow
EnterpriseRAG-Bench leaderboard (Hugging Face): https://huggingface.co/spaces/onyx-dot-app/EnterpriseRAG-Bench-Leaderboard ; benchmark repository: https://github.com/onyx-dot-app/EnterpriseRAG-Bench
PipesHub repository and README (Apache 2.0): https://github.com/pipeshub-ai/pipeshub-ai
PipesHub docs: system overview https://docs.pipeshub.com/system-overview ; deployment https://docs.pipeshub.com/deployment/overview ; private cloud and air-gapped https://docs.pipeshub.com/deployment/vendor/private ; MCP server https://docs.pipeshub.com/mcp/overview ; connectors https://docs.pipeshub.com/connectors/overview








