Navigating the Chaos

Every organization gets its own isolated Cove, enforced at the database level.

Postgres Row-Level Security scopes every query to your tenant, so your data never crosses into another client's Cove and is never used to train public AI models. Access runs on a zero-trust model with credentials loaded from a secure secret store — never hard-coded — and the architecture is built for FERPA-aligned handling of student records.

No — Thought Cove deploys in minutes with no code to write.

You upload documents through the dashboard and the ingestion pipeline handles parsing, structuring, and embedding automatically. If you do have an IT department, they can integrate Thought Cove through our REST API and Google-based authentication.

We've standardized on Google's Gemini models, running on Vertex AI.

Document understanding and synthesis use Gemini (gemini-2.5-flash) and retrieval uses Google's text-embedding-004 768-dimension vectors. Running a single enterprise-grade stack — rather than juggling vendors — keeps results consistent and auditable, and your content is never used to train these models.

It's built not to. Thought Cove answers only from your own documents.

Every answer is generated through a strict retrieval-augmented pipeline with inline citations tied to specific source pages. If the answer isn't in your data pond, the system tells you the documents don't contain it rather than guessing — and deterministic keyword search lets you prove when something simply isn't in the corpus.

Absolutely.

Small businesses, remote teams, and independent consultants use Coves to isolate client projects, query standard operating procedures, and rapidly navigate dense technical manuals.

Just upload the new version.

Because the new version has different content, it's ingested as a separate document rather than overwriting the old one — so the previous version stays in your corpus and you can query or compare the two side by side. Re-uploading the exact same file is safely recognized as a duplicate and skipped.

Yes. Files are encrypted at rest and in transit.

Your source documents are stored in Google Cloud Storage, where every object is encrypted at rest (AES-256) and transferred over TLS, and the structured data we derive from them is isolated per-institution by database row-level security.

No. Each tenant is fully isolated at the database level.

Row-Level Security guarantees that a query authenticated for one institution can only ever read that institution's rows; data never crosses tenant boundaries. The one exception is opt-in consortium benchmarking, which is shared only through an explicit, read-only aggregate view.

You own your data — export everything, then we permanently delete it.

Before you leave you can export your entire data pond as portable Parquet packages. On cancellation we run a tenant purge that permanently removes your documents and chunks from our database, leaving no residual rows behind.

Yes — Thought Cove exposes a REST API for partners.

Authenticated REST endpoints (secured with a bearer token) let you upload documents, run retrieval queries, and export data ponds programmatically, so you can wire Thought Cove into your existing SIS, ERP, or internal portals. The service publishes an interactive OpenAPI reference you can explore directly.

Most documents process in a few minutes, in the background.

Parsing, page-level vision analysis, semantic chunking, and embedding all run server-side after upload, so you don't have to wait on the page — we email you the moment your document is ready to query. Large catalogs naturally take longer.

Yes — a dedicated vision layer reads complex layouts.

Gemini-powered page analysis extracts tables and multi-column text into clean markdown, and tables are never split mid-structure, so figures and tabular data stay intact and queryable.

Single documents can be up to 2GB.

Files upload directly to secure cloud storage via a resumable transfer, so even very large institutional catalogs come through reliably. If you need to move an entire existing database rather than individual files, our direct replication tool can mirror your tenant's slice of the data.

Every chunk is scored for centrality and tagged by rhetorical role.

We build a structural graph of each document and apply PageRank — blended with the model's own epistemic read — to give each passage a centrality weight, then label its argumentative function (claim, data, warrant, rebuttal, and so on). Retrieval can then favor the passages that actually carry your institution's commitments.

Yes — references like 'see Section 4.2' are captured as links.

During ingestion we extract course codes and section references as explicit edges between chunks, so downstream tools can deterministically pull a passage and the section it points to without an LLM guessing the connection.

Yes. Your knowledge exports as an open Portable Knowledge Package.

Each export is an Apache Parquet file plus a manifest.json describing the schema, embedding model, and content hash — a portable, vendor-neutral format that any downstream tool can load directly.

PDFs, Word docs, Excel and CSV spreadsheets, and plain-text or Markdown files.

Thought Cove converts raw institutional documents — catalogs, handbooks, policy manuals, accreditation evidence — into structured, AI-ready data ponds, whatever the original format (PDF, DOCX, XLSX, CSV, TXT, or MD).

Ingestion is billed in normalized 'pages,' checked before each run.

A 'page' is a billing unit normalized across formats: a PDF is billed by its actual page count (any tables or images on a page are included free), a Word document at roughly 300 words per page, a spreadsheet at about 50 rows per page, and a plain-text or Markdown file at about 2,000 characters per page. So a 40-page PDF costs 40 pages, while a 5,000-row spreadsheet costs about 100. Credits are tracked per tenant, topped up via Stripe, and your balance is checked before each run.