Skip to main content

Data flow

Read the figure left to right. Public sources on the left are retrieved read-only and stored in Kumkuat's US-hosted workspace store. Your users interact through the browser. Model providers on the right receive prompts and return text; none of them train on the content, and Kumkuat trains no models of its own.

Data flow diagram: public sources and your organisation on the left, the Kumkuat platform in the middle with web application, REST API, background workers, PostgreSQL, Cloud Storage, Secret Manager and identity, and model providers and output destinations on the right
Figure 1. Green paths carry public content retrieved read-only; blue carries your users' sessions and the content they submit for testing; the dashed purple path is the optional document-library connector, off by default. Workspace data is stored in the United States. Model providers receive prompts and return text.

Table 1 · What is retrieved, from where, and where it lands​

SourceHow it is retrievedFields storedStored inSent to a model?
LinkedIn profiles & postsBright Data dataset API (third-party retrieval partner). Public profile URLs and posts only.Profile: name, headline, about, location, follower count, industry, experience, education, skills. Post: author, date, text, like / comment / share counts, post URL.PostgreSQLYes, as persona context and for tagging
X (Twitter)Bright Data dataset API. Public handles only.Handle, display name, bio, follower count. Post: date, text, like / repost / reply counts, permalink.PostgreSQLYes
Truth SocialTruth Social's public Mastodon-compatible API, requested through Firecrawl. Reposts are skipped.Display name, handle, date, post text, like / re-truth / reply counts, permalink.PostgreSQLYes
BlueskyBluesky's public AT Protocol API, no credential.Handle, display name, bio, follower / post counts; post text, date, permalink.PostgreSQLYes
Publications, newsrooms, RSSKumkuat's own crawler over HTTPS; Firecrawl as a fallback for sites that block plain crawlers.Title, article text, URL, publish date, word count, media provenance (owned / earned / paid).PostgreSQLYes
Press releases & earnings callsFinancial data feeds, keyed by company ticker.Title, body text, company, ticker, date, original URL. Transcripts are stored speaker-segmented.PostgreSQL + Cloud Storage text copyYes
Web search (discovery)Google Search via Gemini grounding, DuckDuckGo, Firecrawl search. Used to find newsrooms, social handles and recent coverage for a named person or company.Candidate URLs and handles; for news search, article title, text, URL, date.PostgreSQLYes, the search is itself a model call
Content your users submitTyped or pasted in the browser, or uploaded (PDF, DOCX, PPTX, XLSX, TXT, CSV, MD). Text is extracted on the platform; a signed, time-limited link to the file is passed to Firecrawl only if local extraction fails.Quick reaction test: text is extracted for that request and neither the document nor its reactions are stored. Narrative tests, chat attachments and knowledge-base uploads: original file, extracted text, title, upload date, submitting user, and the outputs generated from it.Quick test: not stored. Otherwise Cloud Storage (file) + PostgreSQL (text, outputs)Yes, as the subject of the prompt
Persona & audience definitionsCreated by your users, or installed from Kumkuat's catalogue of public-figure audiences and then adapted.Name, role, background, stance summaries, linked public sources, the workspace that owns them.PostgreSQLYes, as persona context
SharePoint / OneDrive / Google Drive libraries (optional connector)Microsoft Graph or Google Drive API with delegated OAuth granted by your user; read-only scopes; tokens encrypted at rest; revocable by you at any time.Selected library or folder: file name, file, extracted text, modified date. Change detection syncs new, updated and deleted files.Cloud Storage (file) + PostgreSQL (text)Only if enabled for that library; enrichment can be left off

What a model provider receives​

A prompt assembled on the platform: the selected persona's profile (built from public sources), the most relevant retrieved excerpts, and the text or question your user submitted. The provider returns text. Calls are made over TLS from Kumkuat's Google Cloud project using Kumkuat-managed provider accounts, or the workspace's own provider key where one has been added under Models & provider keys (see Models and providers). Google Gemini is used on the paid API tier and OpenAI through its API; neither trains on API content. No zero-data-retention addendum is in place today; standard provider API retention applies.

Does a persona query go to a public model?​

Yes, in this sense: when a user asks a persona a question in chat, that question is sent to the configured model (Google Gemini by default), and Gemini's Google Search grounding is on by default, so the model may run web searches derived from the question. The user's text is not used to train the model. Kumkuat can switch web grounding off for your workspace on request, in which case the model answers only from persona context and stored documents.

What third parties never receive​

  • Bright Data and Firecrawl receive handles and URLs to fetch. They do not receive content you submit or any output.
  • Financial data feeds receive company tickers only.
  • No Kumkuat feature posts to a social network, press wire or any external system on your behalf.

Connected document libraries​

The SharePoint and Google Drive connectors exist in the product but are off by default. When one is enabled, the library you connect can be kept out of automated tagging and embedding so proprietary text is not sent to a model provider unless a user deliberately submits it for testing.