How OpenSI processes data
Version 2026-10-28.1. Operator: CHOTA LLC (ООО «ЧОТА»), taxpayer number 690877355, Republic of Belarus, Logoysk, 11 Timchuka Street, room 49. Personal-data contact: n.ogorodniy@chota.by; /support in @chotaGPTbot.
Protection before calling AI
Before sending text to external AI and search services, OpenSI hides recognised phone numbers, email addresses, certain identity and payment details, explicitly labelled names/addresses and secrets on the server. This applies to the question, conversation history, book passages, embeddings and reranking. It is automatic pattern-based filtering, not guaranteed anonymisation: unusual formats, typos or context may go unrecognised.
Do not send passwords, keys, identity documents, health information about an identifiable individual, other people’s personal data or confidential documents. Books are not rejected in full because they contain contact details: detected information is hidden in the passages sent out. Original text passages are stored locally in encrypted form. New personal notes and reminders are not created.
Voice messages and audio are unsupported: the bot does not download or transcribe them. After separate confirmation, a photo is sent to Google through OpenRouter for OCR: Gemini 2.5 Flash-Lite, with approved Gemma routes as fallback. Before sending, the server checks size and format, reduces resolution and removes metadata. Visible personal information is not masked beforehand; do not send private or confidential photos. Recognised text is filtered before further AI processing. Telegram itself receives messages and files sent in chat under its own rules.
Information the service needs
- Telegram user and chat IDs for replies, access control and abuse prevention. IDs are encrypted; HMAC is used for matching. Names and usernames are not saved in the bot profile or included in AI requests.
- Research questions and answers, uploaded book passages, embeddings and saved results, processed only for the selected feature.
- Files and confirmed photos are temporarily processed in server memory. The database stores necessary book passages and the encrypted technical file_id of a pending task, not the original photo. Unconfirmed photo requests expire after ten minutes; the photo is not downloaded before confirmation. Cancellation and data deletion remove pending requests. Audio is not downloaded. Photos reach the external OCR provider; extracted text is filtered before being returned to the user.
- Consent version, time and acceptance/withdrawal action; confirmation of being 18+ and located in Belarus; activation, usage and purchase events without question text.
- Plan, duration, Stars amount, Telegram payment and refund IDs, costs and quota accounting. We do not receive bank card numbers.
Ordinary data necessary for the contract is used to form and perform it, provide access, support and accounting. Separate active consent applies to AI processing of submitted content and the international transfers described here. Providing content is voluntary; without that consent, external AI processing does not start. Service enquiries remain available without AI.
Recipients and destinations
Telegram provides chat and Stars payments under its policy. AI processing uses OpenRouter, Inc. (United States), which routes necessary materials to the selected provider. Depending on availability and the feature, these may be NVIDIA, Google, DeepInfra, CoreWeave, Together AI, OpenAI, Microsoft Azure, Baseten, Perplexity or Voyage AI. These organisations and infrastructure are outside Belarus, including in the United States. Their infrastructure and authorised processors may be in other countries. We do not promise that all operations stay in one country.
External providers receive the question, limited conversation context and necessary source passages after local filtering. Undetected information can remain in the request. Photos are sent only for separately confirmed OCR. Audio, Telegram IDs and usernames are not sent. Search providers receive a search query. Arbitrary websites do not receive the full conversation through the bot. Following a link subjects you to that website’s policy.
Ordinary routes use an approved-provider list and prohibit training on data where the API supports that guarantee. Experimental free NVIDIA routes have a separate test regime and may involve the provider retaining requests to improve its service. They are only for public de-identified educational/test materials. User consent does not change external API terms or guarantee commercial-use compliance. Do not use these routes for any private or confidential information.
The United States and other processing countries may not provide the same protection as Belarus. Data-subject rights, government access, appeals and deletion procedures can differ. OpenRouter generally does not disclose the precise processing location of each request; the operator does not promise full control over copies abroad. Provider information: OpenRouter, NVIDIA, Google, Perplexity, Voyage AI. The operator separately records enabled routes; a material new processing method requires document review and, where necessary, fresh consent.
Retention and deletion
Conversation history is retained for up to 30 days. Books, indexes and saved results are kept until you delete them or after 90 days of inactivity, with a bot warning seven days beforehand. If delivery fails because you blocked the bot, the attempted warning still counts and cleanup proceeds. Legacy notes/reminders are not used for AI answers or notifications. They are included in /export and deleted with content through /delete or inactivity cleanup. Unprocessed jobs are deleted when consent is withdrawn; completed technical jobs containing content are deleted within a day. Audio is not downloaded. Original photos are processed in memory, without permanent storage, and released after recognition or an error.
The /delete command with confirmation deletes content from the working database and withdraws consent. A request already sent to a provider cannot be withdrawn retrospectively. Copies held by Telegram, providers and recipients are governed by their terms and applicable law; immediate deletion of every external copy is not promised. Our backups expire within 30 days; restored data is cleaned again for previously deleted content.
Minimal payment, consent, support and quota records may remain after content deletion for accounting, rights protection, mandatory duties and preventing repeat use of free quotas. Consent evidence and encrypted payment IDs are kept in the bot for up to five years; content-free usage events for up to 90 days. The link between a deleted account and its Telegram ID is kept for up to one year after last activity, then removed; aggregate cost accounting without that link may remain longer. These are operator-defined periods for this service, not a claim that the law sets one period for all records. Separate mandatory accounting documents are retained in company records for the periods applicable to them. These records are not used to restore deleted conversations or for advertising.
Security and your rights
Content, embeddings, Telegram IDs and payment IDs are encrypted, with keys separated from the database. Staff access is limited to support and operations needs. Message text, audio and secrets are excluded from application technical logs. API transmission uses TLS. The model has no access to keys, the user database, a shell or arbitrary command execution. Neither these measures nor the system prompt promises absolute security.
You may obtain processing information, request correction, cessation and deletion where grounds exist, withdraw consent and challenge the operator’s actions. /export provides saved user content. For a full data request, contact the operator or support; an enquiry from the relevant Telegram account suffices to verify ownership — do not send a passport to the bot. The operator responds within the periods set by Belarus Law No. 99-Z, generally five working or fifteen calendar days depending on the request. If another lawful basis requires continued processing, the operator explains the basis and retained information. You may contact the National Personal Data Protection Centre of Belarus.
This policy applies to OpenSI, not all CHOTA services. Changes are published with a new version; changes in processing scope or purposes require necessary fresh consent before continuing.
Scientific library and Mini App
Additional stored data includes folders, source mode, questions, retrieved bibliography, quotations, comparison tables and save markers. Original uploaded PDFs are encrypted in a private database for page viewing. Ownership is checked. Mini App sign-in verifies Telegram’s signature; a secure HttpOnly-cookie session lasts eight hours.
Filtered scientific queries are sent to OpenAlex (OurResearch), Crossref and Europe PMC (EMBL-EBI). They receive the search string and server request network data, not Telegram IDs or automatically added private-library text. This is external processing: https://openalex.org/OpenAlex_privacy_policy.pdf, https://www.crossref.org/operations-and-sustainability/privacy/, https://www.ebi.ac.uk/data-protection/privacy-notice/europe-pmc-website . The app does not determine each request’s precise location. Library mode does not call these catalogues, but necessary filtered passages are sent to AI providers to answer.
Unsaved research history is deleted after 30 days; saved results, folders and documents remain until deletion or the inactivity cleanup described above. Incomplete uploads are removed within a day. Deleting a document also deletes saved private-library answers that could contain its passages. /delete removes all this content; backup expiry remains unchanged. Saving public sources does not give you ownership of them; respect access rights.
Scientific search and additional quotation checks
Publication search uses OpenAlex, Crossref, Europe PMC and official PubMed/NCBI (United States). Additional web search uses Tavily and Perplexity Search. They receive a filtered search string, without Telegram IDs, conversation history or private-library passages. Library mode does not call external catalogues or web search. A result may contain only an abstract or search snippet, labelled in the interface; full-text access is not implied.
When additional semantic checking is enabled, TypeSafe AI, Inc. receives a filtered claim, quotation and limited source context, including a selected private-document passage in Library mode. Telegram IDs, usernames, original PDFs and the full conversation are not sent for this check. This is a separate processor outside Belarus; its policy provides for international processing and does not guarantee storage only on our server. TypeSafe policy, Tavily policy. The check assesses correspondence with the text; it does not guarantee scientific truth, absence of errors or suitability for medical use.
These integrations require current consent. If additional checking is unavailable, the bot reports this. Search results and source passages are treated as data, not instructions to execute actions.
Original supported DOCX, ODT, EPUB, FB2, TXT and MD files are encrypted and deleted under the same rules as PDFs. Text is extracted locally. External AI receives only necessary filtered text passages, not original files.
Website sign-in
Website sign-in is confirmed through your Telegram account. Requests last five minutes; we store a random ID, the hash of a browser secret, a comparison code and a confirmation marker. The secret is held in an HttpOnly cookie. A signed-in session lasts eight hours; the database stores a token hash. Expired requests and sessions are removed during database maintenance. To rate-limit sign-in attempts, a keyed IP-address hash is briefly used; the original IP is not stored for that purpose. Use “Sign out” to end the session; use the bot to delete content and withdraw consent. Web chat and bot share one account and library. Subscriptions are purchased in Telegram.
Scientific catalogues, YouTube and subtitles
arXiv and DataCite receive filtered server-side search queries, not automatically added private-library passages or Telegram IDs. Policies: arXiv, DataCite. CyberLeninka and publisher search uses the previously named search providers; restricted interfaces and paywalls are not bypassed. Google Scholar and CyberLeninka buttons open a query on their websites at your request; automatic retrieval of Google Scholar results is not currently performed.
Video search and current counters use YouTube Data API (Google). Google receives a filtered query or video ID, the server network address and request metadata. This does not sign into your YouTube account or access your email or channel controls. The time counters were retrieved is saved because values can change. Google policy, YouTube terms.
A direct summary request sends the public link to Supadata for existing subtitles, without conversation history, Telegram IDs or private-library passages. Generating new subtitles from audio and automatic paid fallback are disabled. Supadata policy. Returned text is filtered before being sent to the previously named AI providers for summarisation. Footage is not analysed; only retrieved subtitles, sometimes a limited part labelled in the answer, are available.
Public subtitles are cached in server memory for up to 15 minutes. Saved YouTube information and summaries are removed from results within 30 days, including saved answers; other scientific materials follow ordinary retention periods. A repeat request retrieves counters again. Account deletion removes saved results and pending jobs.
Free-plan quotation checking
On Free, additional semantic checking uses Mammoth 1 or Elephant 1 hosted by Baseten Labs, Inc. (United States), through OpenRouter. They receive only filtered claims, quotations and limited source context, including selected private-library passages. Telegram IDs, usernames, the full conversation and original file are excluded. The route requires training-use prohibition in provider settings; this does not promise that external services retain no technical records. Baseten terms and processing information. Paid plans use TypeSafe Jev as described above. If additional checking is unavailable, the answer states this; code still checks the quotation’s presence in the retrieved source.
Ratings and voluntary improvement
👍/👎 saves a rating for the particular version of your answer and the selected reason. You may change or delete it. It is retained for up to 90 days, or less if the associated answer is deleted. Conversation examples are not collected for improvement by default.
With separate voluntary consent, new filtered questions, answers and public-source bibliographies are copied into an encrypted quality-analysis collection for up to 90 days per example. Declining does not limit service access. Private-library answers, combined answers and conversations with private documents in their history are excluded. Filtering does not guarantee anonymity. The collection is not sent to external model providers for training. Withdrawal stops collection and deletes examples; backups expire within 30 days.
Technical records contain operation IDs, model versions, token counts, costs, quotas, latency and refusal codes, without question/answer text, files or internal reasoning. These measurements support operation and spending controls; they are not independent evidence of scientific accuracy.
Additional open sources
OSTI receives filtered search strings for scientific and technical materials; DOAB receives public book IDs for bibliography retrieval; World Bank receives indicator and country codes and periods for observations. They receive the server network address, not Telegram IDs or automatically added private-library passages. Initial DOAB-book and ERIC searches use the previously named search providers. Catalogue availability does not mean every record has accessible or read full text.