Enterprise-Scale LLM Deployment with Licensed Content for Business Intelligence

factiva · Media & Entertainment · 2023 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Licensed business-intelligence search and summarization

Factiva needed to add generative AI summarization and natural-language querying to a decades-old licensed business-intelligence corpus without violating publisher rights, weakening content attribution, increasing hallucination risk, or sending unlicensed content into the generation layer.

Constraints:

Generative summaries could only use content explicitly licensed for GenAI use.
Publisher relationships required security assurances, attribution, royalty flows, and education about AI terms.
The corpus spans thousands of sources, many countries, languages, and nearly three billion articles.
Summaries had to preserve relevancy, recency, context, citations, and content integrity.
Dow Jones needed secure infrastructure for indexing and summarizing a private licensed corpus.

Architecture

Factiva built Smart Summaries as a closed licensed-content summarization system on Google Gemini and Google Cloud. The platform uses semantic search over Factiva's permissioned corpus, applies relevancy/recency/context ranking, feeds only GenAI-licensed content into the summarization layer, returns answers with relevant sources, and preserves royalty, citation, audit, and publisher-permission flows. The architecture treats generation as a language task over retrieved licensed material rather than an open-ended knowledge task.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  publisher-licensing-program["Publisher licensing program"] --> licensed-content-corpus["Licensed content corpus"]
  licensed-content-corpus --> semantic-search-layer["Semantic search layer"]
  licensed-content-corpus --> citation-and-royalty-audit-flow["Citation and royalty audit flow"]
  semantic-search-layer --> google-gemini-summarization["Google Gemini summarization"]
  google-gemini-summarization --> smart-summaries-interface["Smart Summaries interface"]
  smart-summaries-interface --> citation-and-royalty-audit-flow

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Publisher licensing program`	Secures GenAI rights publisher by publisher and title by title.	`Licensed content corpus`
`Licensed content corpus`	Permissioned Factiva archive of publisher and Dow Jones content available for search and summarization.	`Semantic search layer`, `Citation and royalty audit flow`
`Semantic search layer`	Retrieves relevant, recent, contextual content for user queries.	`Google Gemini summarization`
`Google Gemini summarization`	Generates Smart Summaries from the closed licensed corpus on Google Cloud.	`Smart Summaries interface`
`Smart Summaries interface`	User-facing natural-language query and answer surface with key points and relevant sources.	`Citation and royalty audit flow`
`Citation and royalty audit flow`	Tracks content contribution, attribution, publisher permissions, and royalty streams.	—

Data flow

Publishers grant GenAI rights title by title, Factiva marks which content is permissioned, semantic search retrieves relevant licensed material by recency and context, Gemini on Google Cloud summarizes that closed corpus, and citation/royalty/audit infrastructure tracks which publisher content contributed to the answer.

“only the Gen content licensed for Gen is going to be within our Genai tools whether they are our tools or our data capabilities that we'll be powering with our gen content”transcript, stated

“we definitely have a full and robust um citation structure and a royalty and audit flow that has been part of the system since our Inception”transcript, stated

Why this architecture

Factiva chose a closed, licensed RAG-like architecture because its core asset is trusted licensed content and publisher relationships. Restricting generation to permissioned content reduces copyright and hallucination risk, while Google Cloud/Gemini provided a secure partner for indexing and summarizing a private corpus at very large scale.

“ensuring that we would be within a closed ecosystem if you will that there would not be any other content coming in that would wrap around or be making additional kind of assumptions”transcript, stated

“we selected uh Google Gemini on Google Cloud for our uh enhanced semantic search which which is at the foundation of all the summarization tools and we selected them primarily for their security”transcript, stated

Models used: Google Gemini · Google Cloud · Semantic search · Factiva Smart Summaries

Scale

Factiva content spans 200 countries.
“we've got content from uh 200 different countries we uh provide information now in 32 or 33 languages”transcript, stated
The global content-source count is well into five digits.
“we are actually in well into uh the five digits for our content sources globally”transcript, stated
Factiva asked Google to index a corpus described as three billion articles.
“what Google said when factiva told it it wanted to index three billion articles”transcript, stated
Smart Summaries launched with almost 4,000 GenAI-rights publishers.
“when we uh launched our smart summary in November uh we had almost 4,000 Publishers that were part of our geni rights”transcript, stated
The GenAI-rights corpus had grown from 2,000 sources to nearly 5,000 sources.
“rapidly grown from just 2,000 sources um not even six months prior so we saw a very significant increase in velocity we're now nearing 5,000 sources”transcript, stated

Alternatives considered (roads not taken)

Use unlicensed or open internet content around publisher material.

Rejected because: Factiva wanted generation constrained to explicitly permissioned content with no outside material wrapping around the summaries.

“the content that is permissioned for Gen is the content that is delivered in our summarizations and that's always going to continue”transcript, stated

Treat hallucination mitigation as pure model behavior.

Rejected because: Factiva framed the task as summarizing retrieved licensed content with relevancy, recency, and context rather than drawing from broad model knowledge.

“only uses the Gen licensed sources that we bring into it so it is a it is a very closed ecosystem in that standpoint”transcript, stated

Choose a provider without detailed security transparency.

Rejected because: Google showed the network architecture and security design needed for publisher and Dow Jones requirements.

“they showed us a network architecture diagram they showed us exactly how we would be um having our uh our infrastructure built in Secure”transcript, stated

Patterns

`rag-retrieval` (retrieval)

Smart Summaries retrieves only GenAI-licensed Factiva content before summarization.

“this is also a search summarization tool that only uses the Gen licensed sources that we bring into it”transcript, stated

`retrieval-ranking` (retrieval quality)

Factiva's semantic search ranks source material by relevance, recency, and context before summarization.

“they're based off of relevancy recency and context really those core drivers of semantic search so we really wanted to have true search”transcript, stated

`privacy-preserving-processing` (security)

The system uses a closed ecosystem so licensed publisher content is not sent outside the approved summarization path or combined with unlicensed material.

“there would not be any other content coming in that would wrap around or be making additional kind of assumptions within a summarization”transcript, stated

`safety-policy-enforcement` (rights management)

Publisher permissions determine which content can appear in GenAI tools and which royalty stream applies.

“we know exactly who is permissioned for Gen content and who is not permissioned for Gen content and we manage those royalty flows in separate streams”transcript, stated

Tradeoffs

Constrain generation to licensed content.

Chose A closed GenAI-licensed corpus over Open-web or unlicensed augmentation.

Rationale: The closed corpus protects publisher rights, content integrity, and trust while reducing unsupported assumptions.

“the content that is permissioned for Gen is the content that is delivered in our summarizations and that's always going to continue”transcript, stated

Use Google Gemini on Google Cloud for Smart Summaries.

Chose Google Gemini and Google Cloud over An unspecified alternative provider or self-built summarization stack.

Rationale: Factiva selected the provider primarily for security and for the scale needed to process the corpus.

Use best-of-breed infrastructure by need.

Chose Provider choices based on specific use-case fit over A blanket model or vendor preference.

Rationale: The interview frames Google's role as selected for this specific Factiva need rather than as a universal standard.

“I'll just reiterate best of breed solutions that are going to um serve our needs serve our customers and ensure the intellectual privacy and property”transcript, stated

Gotchas (what bit them)

Publisher education became part of the technical rollout.

Root cause: Publishers were confronting unfamiliar concepts such as RAG, hallucinations, agents, and content-security implications.

Resolution: Factiva went publisher by publisher and title by title to secure rights and explain security, content integrity, and compensation.

“Terms like "RAG model," "retrieval augmented generation," "agents," and "hallucinations" were new to the publishing community.”summary, stated

Generative AI can undermine trust if it mixes licensed content with unsupported assumptions.

Root cause: Open-ended generation can wrap source material with outside content or make additional assumptions.

Resolution: Limit Smart Summaries to GenAI-licensed content inside a closed ecosystem.

“ensuring that we would be within a closed ecosystem if you will that there would not be any other content coming in that would wrap around”transcript, stated

AI licensing changes royalty and attribution accounting.

Root cause: Generative summaries create new contribution moments where publisher content must be traced and compensated.

Resolution: Extend Factiva's existing royalty, citation, and audit infrastructure into GenAI flows.

“we're now able to introduce as we have called it new royalty moments uh within generative AI”transcript, stated

Evaluation

Evaluation is described through content-governance and search-quality principles rather than model benchmarks: only licensed content is used, summaries are driven by relevancy/recency/context, and citation/royalty/audit flows track source contribution.

Metrics: Licensed-source eligibility · Relevancy · Recency · Context · Citation and royalty auditability · Publisher permission status

“they're based off of relevancy recency and context really those core drivers of semantic search”transcript, stated

“ensuring our Publishers know exactly how their content is being used where it's being distributed”transcript, stated

Outcomes

Factiva launched Smart Summaries in November.
The GenAI-rights corpus grew from about 2,000 to nearly 5,000 sources.
The system keeps generation within a closed licensed-content ecosystem with citation and royalty accounting.

The case combines a nearly three-billion-article corpus, content across 200 countries, and a GenAI-rights corpus approaching 5,000 sources.

“we uh we just launched that in November so more will be coming”transcript, stated

“we're now nearing 5,000 sources that are part of our gen rights licensed uh content Corpus”transcript, stated