AI Transparency

What Bespoke AI is

Bespoke AI is a personal AI workspace provided by Bespoke Technologies. It combines Bespoke-owned product instructions, routing, memory boundaries, tools, storage, and interface components with configured third-party AI processing.

AI-generated text, images, code, summaries, and other artifacts are generated output, even when the interface presents them in a polished form. They are not statements made or independently verified by a human at Bespoke Technologies.

How intelligence is selected

Bespoke AI routes work by capability rather than promising one fixed model. Auto balances quality, speed, availability, and cost. Fast favours lower latency for everyday work. Expert allocates more capable reasoning routes for complex work. The selected upstream model may change as availability, safety, evaluation, and product quality change.

Before a route is eligible, deterministic controls can check required capabilities, operational health, data-policy classification, regional constraints, context size, and safety limits. Escalation is bounded rather than unlimited.

What context may be used

A response can use your current message, recent conversation, selected attachments, active Project instructions, and memories that you or the product have explicitly placed within the allowed scope. Project-only memory is not intended to cross into another Project or global context.

Submitting sensitive or confidential information is your decision. Use the minimum necessary information and do not submit content you are not authorised to share. See our Privacy Policy for data-handling information.

Memory, and the models it uses

Memory keeps short notes about you so you do not have to repeat yourself. What it holds, how to read or delete it, and what it refuses to keep are described in our Privacy Policy. This section is about the models involved.

Finding relevant notes uses one embedding model: Google Gemini Embedding 2, reached through the Vercel AI Gateway. It converts your question, and each saved note, into a list of numbers that can be compared. Unlike the models that write answers, this one is not interchangeable: notes saved with one embedding model cannot be compared against another, so changing it means rebuilding every saved note, and we would record that change here.

While learning from chats is on, which it is unless you turn it off, a second step reads the message you sent and proposes what might be worth remembering. It runs on the Fast route, which as of this version selects between DeepSeek V4 Flash, Google Gemini 3.5 Flash Lite and OpenAI GPT-5.6 Luna, with Google Gemini 3.6 Flash as a fallback. As with other routes, the selected model can change with availability, evaluation and cost.

In a typed conversation, the model writing your answer can ask to look up your saved notes when it judges they would change the answer. It supplies only a few words to look for; which notes it may see is decided by the product, by the chat's scope and your settings, and the words are converted by the same embedding model. Preferences and constraints you saved or stated reach that model with every answer, labelled as standing preferences and as data rather than instructions.

What that model proposes is not what gets saved. It can only suggest; the product decides, and the decisions are fixed rules rather than model judgement: something you stated may be saved, something merely inferred is either kept as a low-confidence preference or held until you accept it, refused categories are dropped whatever the model says, and a new note never silently replaces one you confirmed. Saved notes reach the model that writes your answer as labelled data, not as instructions, and a note that turns out to be wrong is evidence to be corrected rather than a rule the assistant must obey.

Manual workspace Search is different from asking Bespoke AI to reference an earlier chat. Eligible Projects and chats are found through Bespoke-owned wording search. For Library files with existing indexed passages, the query may be converted to a vector by the configured embedding model through the AI gateway so passages can also be found by meaning. The file passages and chat messages are not sent for a new embedding by this manual search. A result is not placed into an answer unless another separately enabled feature allows that context.

If you turn on searching your earlier chats, excerpts of those conversations are converted by the same embedding model, Google Gemini Embedding 2 through the Vercel AI Gateway, and changing that model would mean rebuilding them as well. No model decides what is kept or what is found: excerpts are cut from each exchange by fixed rules, a search happens only when fixed wording rules find that your message refers to an earlier conversation, and which excerpts reach an answer is decided by ranking rather than by model judgement.

Excerpts that reach an answer are given to the model that writes it labelled as earlier conversation, with the date they were said. That model is told to treat them as information rather than instructions, and as weaker evidence than a saved note, because a note is something you chose to keep and an earlier conversation is only something that was said.

Live response states and reasoning

The interface may show measured activity states such as preparing, thinking, searching, or writing while a response streams. These labels describe the current product phase. They do not guarantee a particular method, level of effort, or successful result.

When a reasoning disclosure is available, it is presented as supporting context for the response. It may be condensed, incomplete, or generated for explanation and should not be treated as a complete record of every internal computation.

Tools, search, and sources

Some capabilities may use Bespoke-owned tool boundaries for search, files, projects, or artifacts. The interface should show a truthful activity or result when a tool is used. A citation means a source was associated with a claim or result; it does not mean Bespoke Technologies endorses the source or guarantees its accuracy.

Open cited material and compare it with the response, especially for recent, technical, legal, medical, financial, or safety-critical information. A response without citations should not be assumed to have been checked against current sources.

Known limitations

• Output may be inaccurate, incomplete, outdated, biased, offensive, or internally inconsistent.
• The system may misunderstand instructions, images, audio, files, local context, or the significance of missing information.
• Generated code may be insecure or fail outside the shown example. Test and review it before use.
• Generated output may resemble existing material and may not be unique or suitable for registration, publication, or commercial use.

Professional and high-impact use

Do not use Bespoke AI as the sole basis for a decision that materially affects a person or as a substitute for a qualified professional. Apply human review, domain expertise, testing, and the safeguards required by law and your organisation.

AI providers and processing terms

Bespoke AI may use more than one AI provider through a governed gateway or a capability-specific adapter. Provider identities are implementation details and are not a quality guarantee. Provider-specific retention, location, and model-improvement terms can differ; our Subprocessor List records material processing relationships as they are approved for publication.

Feedback and concerns

Report a harmful, inaccurate, or unexpected result through the Bespoke Technologies contact page. Do not include more private information than is needed to investigate.

Changes and version history

We update this notice when a material AI role, routing rule, data-use boundary, or user safeguard changes. Material changes receive a new calendar version and preserve the prior notice.