TrustCore logoTrustCore
Answer Questions About Your Own Documents, On Your Own Hardware

Your documents stay on your machine. Every answer cites its source.Built for legal, compliance, and investigation work, where the documents are often privileged, confidential, or personal data under the GDPR.

Contact Us
See Pricing

What TrustCore Does

Answers You Can Trace, Without Sending Your Documents Anywhere

Two things are always true: your documents and questions stay on your machine, and every answer is tied to the exact passage it came from.

  • Runs Where You Already Work

    A panel on ChatGPT, Gemini, Claude, DeepSeek, Perplexity, or a local OpenWebUI. The chat site keeps working as normal; TrustCore does not reroute the page's own input box.

  • No SaaS, No Backend

    There is no hosted service to sign in to. The extension runs in your browser; the optional companion runs in a container on your own machine or server. Nothing is uploaded to a TrustCore server.

  • Graph RAG, Not Similarity Alone

    Retrieval combines vector search, full-text BM25, and rank fusion, then walks a graph of the entities in your documents to pull in connected facts a similarity search would miss.

  • Every Answer Cites Its Source

    Each cited quote is checked against the source text and page before the answer is shown. In the strict setting, an answer whose citations fail the check is withheld.

  • Cross-Document Conflict Detection

    When two source documents state incompatible values — two termination notice periods, two amounts for the same item — TrustCore names both sources and pages instead of silently picking one.

  • Local Embeddings on Your GPU

    The bundled bge-m3 model turns your text into vectors on your GPU through WebGPU, with a CPU fallback when there is no usable GPU. On the local path, no embedding call leaves the machine.

  • Located Down to the Page

    Text-based PDFs are read directly; scanned PDFs are recovered by OCR; plain text and Markdown are read as-is. Scanned PDFs are parsed by MinerU in the local companion, which returns text plus page and bounding box per block — what lets TrustCore validate a quote to its exact location on the page.

  • Vision Fallback

    When the companion is not available, TrustCore falls back to a per-page vision model, which gives page-level location instead of block-level.

  • Nothing Read Silently

    A file in a format TrustCore does not recognize is rejected rather than indexed partially, so a project never contains a document that was only half read.

Sovereignty & Compute

Sovereign by Default, Flexible on Demand

Using It Does Not Create a New Place Where the Documents Live

For privileged, confidential, or GDPR-regulated material — and work under professional secrecy such as secreto profesional in Spain or solicitor–client privilege in Ireland — uploading documents to a shared third-party service can itself be the problem. And when you connect external compute for heavier models, the rest of the system behaves the same — only where the numeric work runs changes; the only processor involved is the endpoint you control.

No third party in the data path

On the local path there is no cloud host, analytics service, or model API processing your data. A party that is not in the path cannot leak a document or personal data.

On-device storage

Documents, text chunks, embeddings, the full-text index, and the entity graph are kept in a SQLite database inside the browser, persisted to the Origin Private File System. There is no sync service behind it.

You choose what stays local

Questions asked in TrustCore's panel are answered on your machine. Anything you type into the chat site's own box still goes to that site as usual. The boundary is explicit and under your control.

OpenAI-compatible interface

All chat and embedding calls use the OpenAI-compatible request format, so TrustCore can talk to OpenAI, OpenRouter, a local Ollama or llama.cpp server, or any on-prem endpoint that speaks the same format.

Point at infrastructure you control

The paid tiers unlock pointing TrustCore at your own embedding, OCR, and chat endpoints — on-prem or in your own cloud account — so heavier models run on infrastructure you control.

The data boundary stays yours

With the paid path the only processor involved is the endpoint you select and control. It is your subprocessor under your contract, not one TrustCore introduces.

Retrieval & Reasoning

Graph RAG, Not Just Similarity

Graph RAG retrieval over document entitiesTrustCore builds a graph of the people, organizations, dates, amounts, and references in your documents and reasons over it, so connections invisible in any single file become queryable — and for work where a wrong or unsupported statement has consequences, the evidence arrives attached and already checked.

Hybrid retrieval

Vector search finds passages close in meaning; BM25 finds exact words that matter for defined terms, clause numbers, and names; rank fusion merges the two into one ranked set.

Graph expansion

Starting from the entities in the top passages, TrustCore walks the relationship graph up to several hops to bring in connected facts across documents.

Why it matters here

In legal and compliance documents, similar wording appears in a master agreement, a superseding amendment, and a schedule. Similarity alone can answer with high confidence from the wrong one.

Proof paths

A symbolic reasoner produces an explicit chain of relationships from the question's subject to the supporting facts, so a person can see why the answer follows.

Citation admissibility check

Before an answer is shown, each cited quote is confirmed to appear on the page it is attributed to. In the strict setting, answers that fail are withheld rather than shown.

Temporal and superseded-clause reasoning

For a chosen "as of" date, the reasoner resolves which provision is in force, labels superseded clauses with their effective dates, and shows both — never silently dropping the old one. Available on the Team plan and above.

The Graph You Can See

See, Check, and Correct the Graph Your Answers Are Built From

Built Read-Only From On-Device Data, Drawn on Your GPU

The map renders through WebGPU, with WebGL and a 2D view as fallbacks. Nothing is sent anywhere and building it makes no model or companion call. Very large graphs are capped at the most-connected nodes, and the map states when it has done so.

See why an answer holds

The map shows the same entities and relationships an answer is built from. Selecting any node opens the exact source passage and its page, so a reviewer audits the evidence by eye — opened per project from the ⬡ 3D Map button in the panel.

Connections no single document states

When entities are linked only through a third file, a similarity search returns nothing; the graph records the link and the map shows it. Focusing a node lights up its neighbours across files, community detection colours related clusters, and the most-connected entities are drawn larger.

Correct it, then export it

Mark a clause superseded by another by hand; that link is treated as confirmed, survives re-indexing, and feeds the point-in-time reasoning. The whole map exports as an Obsidian-compatible vault — a ZIP of linked notes — for a case file.

Pricing

Fixed Annual Packages for Teams of Every Size

Start fully local at no cost. Paid tiers add the capability an organization with its own infrastructure cannot get for free: connecting your own compute. One-year licences for 1, 10, or 50 seats — prices are ex-VAT.

Free

Fully local, single user

Free
per user
  • 100% local — bundled embeddings, OCR, and model
  • Chat-site integration and validated citations
  • Hybrid retrieval (vector + BM25 + rank fusion)
  • Capped projects and documents
  • No external compute, no subprocessor
Contact Us
most popular

Starter

1 seat

€348
per year+ 21% VAT at checkout
  • Everything in Free
  • Connect external compute — on-prem or cloud
  • Cross-document conflict detection and proof paths
  • Full citation verification
  • Unlimited local projects
Buy Now

Team

10 seats

€2880
per year+ 21% VAT at checkout
  • Everything in Starter
  • System Prompt per project
  • Temporal / superseded-clause reasoning
  • Shared, centralized provider configuration
  • Priority support
Buy Now

Business

50 seats

€11400
per year+ 21% VAT at checkout
  • Everything in Team
  • Admin policy controls — force endpoints, lock model
  • Onboarding
  • SLA support
Buy Now

FAQs

TrustCore — Frequently Asked Questions

Straight answers for legal, compliance, and investigation teams evaluating where their documents go.

Do my documents get uploaded anywhere?
No. On the local path your documents, the extracted text, embeddings, the entity graph, and your questions stay on your machine. There is no TrustCore server, and document content is not shipped in telemetry. If you choose the paid external-compute path, the only place data goes is the endpoint you select and control.
Which chat sites does TrustCore work with?
TrustCore opens its own panel on ChatGPT, Gemini, Claude, DeepSeek, Perplexity, and a local OpenWebUI. It answers from your indexed documents in that panel. The chat site's own input box keeps working as normal — TrustCore does not intercept or reroute it.
Do I need a GPU?
No, but one helps. TrustCore runs on a Chromium-based browser (Chrome, Edge, or Brave) version 124 or later, on Windows, macOS, or Linux, with 8 GB of memory (16 GB recommended for larger projects). Any GPU that exposes WebGPU accelerates embedding and the graph view; without one, embedding runs on the CPU and is slower.
What document types are supported?
Text-based PDFs are read directly. Scanned, image-only PDFs are recovered by OCR. Plain text (.txt) and Markdown (.md) are read as-is. A file in a format TrustCore does not recognize is rejected rather than indexed partially, so a project never contains a half-read document.
How do I know an answer is correct?
Every answer is tied to the exact passage it came from, and each cited quote is checked against the source text and page before the answer is shown. Where two documents disagree, the conflict is flagged with both sources and pages. A proof path shows the chain of relationships behind the answer, so you can see why it follows rather than trusting that a model produced it.
Can I see and check the graph behind an answer?
Yes. The project's entity graph opens as an interactive 3D map from the panel. Selecting a node opens the exact passage and page it came from, so you audit the evidence by eye. The map renders on your own GPU through WebGPU with nothing sent out. You can mark a clause superseded by hand — that correction survives re-indexing — and export the whole graph as an Obsidian-compatible vault.
What is the difference between Free and the paid tiers?
Free runs entirely on your machine with the bundled local model, with capped projects and documents. The core paid unlock across every paid tier is connecting your own external compute — on-prem or cloud embeddings, OCR, and chat endpoints — together with cross-document conflict detection and proof paths, and unlimited local projects.
What is temporal / superseded-clause reasoning?
Available on the Team plan and above, it answers "as of" a chosen date by resolving which provision is in force from the recorded amendment and effective-date relationships. The decision is computed by the reasoner, not the language model. Superseded clauses are labelled with their effective dates and shown rather than silently dropped, and a reviewer can set a supersession by hand.
Do I need the local companion?
Only for some tasks. The companion is a small container stack that handles scanned-document OCR with block-level provenance, cross-encoder re-ranking, and serving a local model. Everything else — indexing, hybrid retrieval, local embeddings, validated citations — works in the browser without it.

Keep Your Documents Where They Are

Tell us about the documents your team works with and the infrastructure you control. We will show you how TrustCore fits, on your hardware and your terms.

Contact Us