Best RAG data providers for licensed knowledge in 2026

Start with the questions the application must answer. Historical research might need a fixed archive; a current-events assistant needs a way to receive new material. I would use these five checks before comparing catalog size.

1. Match the coverage to the questions

Name the subjects, languages and history window you need. Ask whether the proposed collection contains complete works, excerpts or search results. Owners should identify excluded titles and third-party material alongside what they can supply. Give every supplier the same representative questions and ask it to identify relevant material within the proposed license. Review an authorized sample or source inventory. A total catalog count tells you little if the supplier cannot show where your questions are covered.

2. Define what the product can retrieve and show

Walk through one intended answer, from stored source text and embeddings to the retrieved passage, summary and visible citation. Ask the supplier to map each step to the proposed permission, including retention and attribution. Keep training separate. For owners, this means identifying the rights they hold and those that need another person’s agreement. For buyers, a working API key is only the beginning: an answer about download access does not settle the uses the product requires.

3. Set the freshness and correction requirements

For a fixed archive, agree a version and cut-off date. For changing material, ask how new items, corrections and withdrawals reach you. Inspect one item’s publication, modification and delivery timestamps, then follow a correction into the stored copy. Put the timing and responsible role on each side in writing. Owners need to price that recurring work; buyers need someone who will apply the changes. “Live” is an incomplete answer without a delivery cadence.

4. Check the delivery format and source metadata

A file delivery, a feed and a finished answer give your team different amounts of control. Request a sample in the proposed format and trace a passage back to its source record. Check identifiers, titles, dates and the fields your interface needs for attribution. Record any conversion the buyer must perform and any metadata the owner must maintain. A polished answer demo is insufficient evidence when your application needs source records for its own retrieval and correction process.

5. Compare the cost of the agreed scope

Price the same content, uses, cadence and term at expected usage and a higher workload. Identify what creates a charge: a request, a returned item, a subscription period or a delivered collection. Include preparation, platform charges, overages and renewal. Keep the buyer’s invoice separate from the owner’s proceeds. An owner also needs to know what update or reporting work that payment covers. Ask what changes if usage falls, rises or stops; an unexplained billing unit prevents a meaningful comparison.

These models price different purchases. Compare a written scope before comparing totals. “Not published” below refers to the cited offer; request the terms that apply to your purchase.

Rightmark: best for scoping specialist archives for RAG

We arrange scoped document and archive licenses between owners and AI teams. A specialist publisher’s archive, courseware or community knowledge starts with a specific request: which subjects and years the buyer needs, what the application will show, and which rights the owner holds. Our documents route treats training, retrieval and display as “Separate rights, priced separately.” Ownership stays with the owner. For the buyer, that means specifying a collection and its intended use together. For the owner, it means documenting exclusions early, before a buyer designs answers around material that cannot be included in the proposed archive.

The buyer workflow describes text by title or archive with a rights record per item, a written delivery scope and owner signature before licensing. For news archives, we specify clean text with dates, bylines and sections. Those fields give a buyer the context needed to inspect a passage and plan its attribution. Owners must separate material they control from freelance or wire content. Samples are under NDA for evaluation only; agree the proposed checks and handling before putting them into an application.

On standard licenses, the buyer pays our fee on top of the owner’s payout; premium and bespoke shares are agreed in writing. The book and forum lots shown in our buyer workflow are in preparation, so confirm the requested collection and cadence during scoping. For current news, a direct publisher feed is a stronger fit than our archive-scoping route. Owners can start an appraisal with a description, without uploading files or granting system access. Acceptance and remedy terms remain in development; obtain the terms for your proposed agreement.

Best for buyers and owners scoping a specialist archive for defined RAG use; choose a direct feed supplier when immediately available, continuously updated news is essential.

The Associated Press: best for grounding answers in ongoing news

AP supplies news directly to AI products that need to answer questions about changing events. Its AI, Tech and Data offering covers grounding and display applications alongside training, with ongoing and archival material across subjects including politics, business and sports. Structured content and metadata support ingestion. Start by specifying the reporting your users need: topics, languages, media types and historical depth. A useful news source should answer the questions in your application, including the follow-up that asks what happened before the latest development.

The Media API provides programmatic access to licensed content, including text stories with linked material. Its documentation describes 30 days of content and a fuller archive for pictures, video, audio and print graphics; ask separately about older text. In AP’s January 15, 2025 report of its Google agreement, Google described “a feed of real-time information to help further enhance the usefulness of results displayed in the Gemini app.” That was the announced purpose. Payment and presentation details were undisclosed.

Ask for a proposal covering retrieval, stored copies, displayed excerpts, attribution and correction delivery for your chosen content. An owner negotiating a similar feed should price both the permitted answer use and the obligation to maintain that feed. AP’s cited AI offering publishes no RAG price or minimum commitment. Its news coverage suits a different requirement from a private company’s operating history. A team answering questions about internal procedures would still need its own records. For publishers with unrelated catalogs to license, an owner-facing intermediary is a more relevant starting point than a subscription to another publisher’s news. Include the required text-history window in the proposal itself. That gives the technical team a concrete entitlement to check, instead of discovering during ingestion that the requested period falls outside its access.

Best for answers that depend on ongoing news; teams needing private company operating history should choose an archive specialist.

Linkup: best for enterprise access to licensed AFP news

Linkup’s licensed AFP route gives enterprise teams access to news dispatches through the API they use for retrieval. The companies announced their signed agreement on October 28, 2026. Enterprise partners can query AFP alongside Linkup’s web index. This is a specific publisher-content arrangement for teams that want AFP reporting in an assistant or agent workflow. Define which questions need that reporting and what a satisfactory answer looks like before discussing integration. A source requirement is easier to evaluate when the expected output includes its citation and the passage supporting it.

Linkup describes the premium-source model as licensed, attributed and approved use case by use case. In the announcement, Patrice Monti, AFP’s Sales & Marketing Director, says: “Working with Linkup lets us bring AFP journalism to a new generation of enterprise products”. He speaks for the participating publisher. Ask for a representative response that distinguishes AFP material from wider-web results. Put the attribution, storage and displayed-answer requirements for each source into the proposal, so the shared interface does not obscure different permissions.

Linkup’s general terms say it has no rightsholder license agreements for ordinary open-web content linked through the API. Specific terms take precedence, making the AFP agreement’s scope essential. The pricing page offers custom enterprise pricing; its general Fetch, Search and Research rates do not quote AFP rights. Request content charges and platform charges separately, with the term and permitted use. A publisher considering this model needs compensation and approved uses for its own catalog. The AFP announcement names one participating publisher and supplies no general publisher payout schedule or self-service access route. An owner should also define what happens when it corrects or withdraws an article, and which party maintains the corresponding copy in the application.

Best for enterprise teams seeking licensed AFP news through a retrieval API; buyers needing self-service news rights should look elsewhere.

Cashmere: best for licensed publisher passages with citations

Cashmere connects publisher catalogs and metadata to AI applications through a consistent API. Its product guide identifies books, paid blogs, market research and academic journals, with publishers defining access and AI contexts. Its terms apply licenses to individual items or collections; when publisher and platform terms conflict, the more restrictive terms govern. Choose the collections that cover your questions, then specify the answer use. For a publisher, the corresponding choice is which works to make available and which applications may use them.

The RAG cookbook returns passages with relevance scores, source titles, publishers, section labels and citation links. Developers pass selected passages to their model and preserve the returned citation URL. Cashmere checks access on click and redirects to the publisher’s configured destination. Dave Brooks, identified as Sr. Director Go-to-Market Ops at Wiley in Cashmere’s homepage testimonial, says: “They got our content AI-ready at pace and scale”. Publisher plans are separate from buyer content fees: Control is free, Monetize uses revenue share and Scale is custom. Standard ingestion is included; advanced processing is usage-based.

Citation handling needs attention. The first click starts a 24-hour window; afterward, visitors without their own access see an expired-link page. Collections marked no license required are exempt. Redisplayed answers need a new search for fresh links. Content retention is a separate decision: termination generally requires deletion, except for separately licensed content or publisher terms allowing retention. That distinction matters for a saved research workspace. Decide how old answers will remain useful, and request the appropriate retention permission before treating retrieved passages as a permanent corpus. Owners should include complex file preparation and continuing delivery work in the commercial discussion. For a research product, test reopening a saved answer alongside testing a new query. The same source requirement should cover both experiences before procurement commits to the integration. If Cashmere terminates the platform for convenience, its terms provide a prorated refund of prepaid unused fees.

Best for licensed publisher passages with source citations; teams needing durable public citation links without fresh searches should choose another delivery model.

TollBit: best for publisher-controlled RAG access

TollBit gives publishers infrastructure to set the terms and rates for AI access. Its standard summarization license covers single-use summaries, citations or grounding; a separate standard license covers single-use full-article display. Both exclude training and generative reuse. Publishers choose the license and eligible AI systems, and custom agreements can run alongside the standard options. Start with the actual output: a brief summary and a screen displaying the whole article call for different permissions. An owner can decide which of those uses to offer and price.

AI systems can discover terms before retrieving content, and licensed requests are logged. The developer agreement puts the content license directly between publisher and buyer; TollBit supplies the platform. In TollBit’s September 2025 DataDome integration announcement, Patrick Hainault, VP Corporate Business Development at Mansueto Ventures, publisher of Inc. and Fast Company, says: “TollBit manages access rights to our content by AI providers, based on rates we define, and handles payment collection.” His account concerns that joint deployment.

Buyers pay publisher-set content fees plus applicable portal-listed platform charges. Owners can incur fees for elected additional services. TollBit’s publisher agreement does not guarantee developer payment or compliance with the content license. Ask how collection, reporting and payment timing will work before forecasting receipts. Request logs show access activity; agree how subsequent use will be handled too. A buyer planning repeated reuse or a retained corpus needs a license covering it. Standard single-use access is a different purchase. Put those requirements into the agreement before building storage around retrieved articles; otherwise, the integration may work while the intended reuse remains unresolved. For an owner, a useful pilot report should reconcile the counted requests with the billable events in the agreement, so usage records and expected receipts can be compared.

Best for publishers pricing AI access and buyers seeking scoped retrieval; agree a custom license first if the application needs a reusable corpus.

Created by Humans: best for book reference rights with author approval

Created by Humans connects book rights holders with AI developers and separates training, reference and transformative rights. Build a reference collection title by title: Douglas Preston’s library page marks The Lost Tomb and Extinction with Reference rights, while The Monster of Florence is Unlicensed. The Lost Tomb listing supplies ISBN, publisher and language fields. Use those identifiers in the request and confirm the edition and delivery format. A familiar author’s name is less useful to procurement than a precise list of the works and rights needed for the application.

The author workflow moves from identity verification and claiming books to choosing training or reference preferences, then approving or declining licensing offers. Created by Humans describes curation, file delivery, outbound licensing and payment handling for individual works or bundles. Douglas Preston, a founding author participating as a rights holder, says in the launch post that “creators retain control of their work”. For owners, offer approval is a chance to assess the actual use. For buyers, the workflow coordinates several rights holders without removing the need to specify editions and metadata.

The company offers buyers one agreement and invoice across rights holders. Request a quote for the selected collection, including retrieval, excerpts, summaries, attribution, stored copies and deletion. Owners should ask for the payment basis, deductions and timing for the proposed offer. The Reference label is a useful starting point, but the application’s permitted behavior belongs in the signed scope. This route suits a deliberately selected book collection. A product answering current-event questions needs a separate update service; choose a news-feed provider for that requirement. Confirm ongoing access before incorporating a listed title into the product’s corpus. Use a small set of intended questions to check the proposed books before expanding the collection. A missing title may matter more than the total number offered.

Best for selected books licensed for reference use with owner approval; choose a news-feed provider when current-event coverage drives the application.

Stack Overflow: best for technical Q&A with community context

Stack Data Licensing offers Stack Overflow and Stack Exchange questions and answers as a full corpus or tailored subset. RAG search is an identified use. Tags, comments, votes and revisions can accompany the content, giving a technical assistant more context than isolated answer text. Scope the topics and sites your users need, then test whether an answer still applies to their software version and constraints. A highly voted response can be useful evidence to inspect without becoming an automatic recommendation for every environment.

The documented API returns JSON, supports custom field filters and exposes revisions. Its throttles and caching make update handling part of the integration discussion. In the February 2024 Google Cloud partnership announcement, CEO Thomas Kurian described “combining the vast knowledge from the Stack Overflow community and new AI capabilities”. Google planned to surface suggestions, code and answers through OverflowAPI; Stack Overflow selected Google Cloud to host its platform. For your own product, trace a revised answer through the ingestion process and check what appears in a saved response after that change.

The partnership route emphasizes attribution and prioritizes partners contributing back through two-way integrations. Ask for commercial terms covering the selected corpus, answer and code display, retained copies and deletion. Ordinary API credentials do not establish that agreement. An owner or community operator should also distinguish access to this corpus from a route for licensing a separate archive: identify what each party would contribute before evaluating an integration. Technical Q&A will not answer questions that depend on your company’s private procedures. Keep those internal sources in the retrieval plan, with their own permissions, rather than expecting a public knowledge collection to supply missing operating context. Keep questions attached to answers during that test: the original environment and constraints can explain why a technically plausible answer does not fit your user’s problem.

Best for technical assistants needing licensed Q&A and community context; private operating knowledge needs another source alongside it.

1. Keep the sources that fit the same requirement

Keep 2 to 3 candidates whose proposed content and uses fit the question set. Compare fixed collections with fixed collections, or explain why the application needs a feed. Owners should list the preparation and recurring duties they can accept; buyers should name who will operate ingestion, citations and updates. Copy the blank buyer request brief into an internal document and add one column per candidate. Record its answer, supporting evidence and unresolved items against each requirement.

2. Ask every candidate the same seven questions

Use these questions on each call. The red-flag replies below are invented examples, not statements made by any provider.

1. “Which exact works, subjects, languages and date ranges answer these questions, and what is excluded?” Match an authorized sample or inventory to the same question set. An invented red-flag reply: “Our catalog has millions of items, so your subjects must be covered.”

1. “Who can authorize this content, and which terms cover storage, embeddings, retrieval, displayed excerpts or summaries, attribution and retention?” Trace one intended answer to the offered rights and resolve third-party exclusions. An invented red-flag reply: “You can download it, so any use is covered.”

1. “What version will we receive, and how will new material, corrections and withdrawals reach us?” Name the timing, identifiers and person responsible for updating stored copies. An invented red-flag reply: “The feed is live; you do not need a correction process.”

1. “Can we trace a delivered passage to its source and show an accurate citation?” Inspect the required fields and reopen an answer later to check citation access and expiry. An invented red-flag reply: “The demo looks right, so you do not need source records.”

1. “What does this scope cost at expected and higher usage, and how are owner payments calculated?” Reconcile preparation, platform fees, overages and renewal with the billable unit. Separate buyer spend from owner proceeds and payment timing. An invented red-flag reply: “We can explain the charge after the first invoice.”

1. “Which evaluation activities may we perform, with which people and model services?” Get written permission for sample storage, embeddings, model calls, displayed outputs, logs and deletion before processing. An invented red-flag reply: “The NDA or trial key is enough; send it to any model.”

1. “What counts as an accepted delivery, what happens if tests fail, and what must each side do when the pilot ends?” Agree the failure review, retest or remedy, access closure, retained copies, deletion and outstanding payments. An invented red-flag reply: “Acceptance and deletion can wait until the annual agreement.”

3. Put a 30 to 60 day pilot in writing

Propose a 30 to 60 day evaluation with start and end dates. Agree permitted activities before handling content. If processing is not yet authorized, begin with an inventory or schema discussion and keep the data outside the application.

• Content and workload: Name works or a selection rule, exclusions, languages, version, volume, request limit and a fixed question set. Assign preparation and update duties.
• Permitted use: Specify storage, embeddings, model services, outputs, attribution, recipients and retention. Name who approves expanded use.
• Delivery: Agree formats, identifiers, metadata, timing, corrections, withdrawals and review contacts.
• Cost: Write down the price or formula, included work, spending limit and overage approval. Separate source charges from model and infrastructure costs, and record the owner’s payment basis and trigger.
• Acceptance: Agree numerical targets for question coverage, passage traceability, citation correctness, freshness and cost per tested workload. Every tested item and use must be authorized. Distinguish missing source material from the application’s retrieval or generation errors.
• Failure and exit: Define underperformance and the agreed response: correction, a bounded retest, or a specified credit or refund. On expiry, stop pilot access unless extended in writing. Specify what happens to source copies, embeddings, logs and backups, and settle outstanding payments.

Use the same acceptance method for each candidate and assess the owner’s workload alongside the buyer’s results. A different proposed evaluation can still work if it bounds permission, cost, acceptance and exit. Leaving those decisions until after a long commitment defeats the pilot’s purpose.

Choose the collection whose coverage and permitted uses fit the answers you intend to show. Then compare the work and cost of keeping it useful. Buyers need citations and stored copies they can maintain; owners need defined uses, manageable duties and an agreed payment basis. Narrow any unresolved requirement before committing.

Prepare your source requirement with the buyer request brief