How to get a TypeSafe AI API key

What is the Jev model from TypeSafe AI?

Jev is a decision model from TypeSafe AI, a company in the United States. A decision model is not a chatbot. You give it some text and a few typed questions, and it returns a probability for each answer instead of writing a reply. Besidka asks it only yes/no questions, such as "does this message refer to an earlier conversation?", and compares the probability with a fixed threshold.

That makes Jev a different tool from GPT, Claude or Gemini:

• It cannot write. It cannot answer your messages, summarize a chat or write a memory entry, so it never replaces your chat model or your memory model.
• It is fast. In Besidka's tests, a check with a short message and a few earlier turns came back in about 170 to 200 ms (median), and in about 210 to 310 ms for 95% of calls. Requests of up to about 27,000 tokens still came back within about 320 ms for 95% of calls. Besidka still gives every call a hard time limit: 800 ms while you wait for an answer, 1.5 seconds in the share window and 2.5 seconds for background work.
• It is cheap. TypeSafe AI lists $0.042 per million input tokens, and output is free. A typical check costs about $0.00002. The whole test run behind this page, 1,006 calls, cost $0.21.

Besidka pins one version, jev-1.14.0. The thresholds were tuned on it, so a newer version is tested again before the pin moves.

Why Besidka uses a decision model for smart checks

Some useful decisions in a chat app are too small to ask a chat model about and too fuzzy for a fixed rule. Does "and in Poland?" need a web search? Does a long coding chat hold anything worth remembering about you? Is this paragraph in a project file trying to give the model orders? A chat model could answer, but it would cost more and add seconds before your reply. A word list is free, but it misses everything it does not list.

Jev sits between the two. Besidka's own rules run first and stay in charge; Jev adds a second opinion where a yes/no answer saves you money, time or a mistake. Every one of these smart checks is optional, runs on your own TypeSafe AI key and has its own switch under Profile, API Keys, Smart checks.

The seven smart checks

Each check below says what it does, what Besidka sends to TypeSafe AI, and what happens if TypeSafe AI is slow, unavailable or rate-limited. Never-stored values, such as card numbers, passwords and API keys, are removed from the text before every call.

Memory checks

A second opinion before a memory entry is saved. Jev looks for sensitive information, such as health or political opinions, and data that is never stored, and for entries a model wrote it checks that the entry really comes from your text. It can only add blocks: it can refuse an entry Besidka's built-in checks allowed, never allow one they refused.

• Sent: the entry text and a short excerpt of the text it came from. Entries the built-in checks already refused are never sent.
• If it fails: the entry is saved on the built-in checks alone.

Auto web search

With a search provider picked and Auto on in the chat input's web search menu, Jev decides for each message whether a web search is needed. A link, a domain or a word such as "today" or "latest" always searches without asking Jev. Each answer shows whether Auto searched. Skipping a search you do not need saves the search fee, and Brave and Exa charge a few dollars per 1,000 searches.

• Sent: your message and your last three turns, each cut to 4,000 characters.
• If it fails: Besidka searches, as it does without the key.

Memory pre-filter

Before memory is generated from a chat, Jev checks whether the chat holds anything worth remembering about you. If it does not, the memory model is skipped, which saves a model call on your own key. It runs only for Generate memory from chats, Build memory from past chats and a project Refresh, never for Update memory now or Import memory.

• Sent: the chat text, in pieces of up to about 8,000 tokens. The chat is skipped only when every piece scores low.
• If it fails: memory is generated as usual.

This check sends the most text, because it reads whole chats. Switch it off on its own if that matters to you.

Prompt-injection screening

A prompt injection is text written to give orders to an AI that reads it, hidden in a file or a web page. Jev screens project files and notes when you add or edit them, and Brave and Exa search results before the model reads them. A flagged file is marked "needs review" and stays out of your chats until you choose Allow anyway. A flagged search result is left out, and the model is told how many were withheld. Nothing is ever deleted.

• Sent: the extracted text of project files and notes, in chunks of about 2,000 tokens, and the title and snippet of each Brave or Exa result.
• If it fails: files are added anyway and checked again later, usually within the hour. Search results are used unscreened.
• Not covered: results of a chat model's own built-in web search, which go straight to the model, and anything a PDF holds outside its text layer, such as a scan.

Search re-ranking

When the model searches your past chats or a project's files, Jev puts the most relevant results first. It only reorders what the search found and never adds a result. It can drop clearly irrelevant results, but only when enough relevant ones remain.

• Sent: the search query and up to 30 short excerpts of about 300 tokens each.
• If it fails: the original search order is kept.

Share check

When you open the share window, Besidka checks the chat for personal data before you publish a public link. Its own detectors look for emails, phone numbers, card numbers and similar identifiers first; Jev then reads the rest for things like addresses, health details or a private person's name. It is a warning only and never blocks sharing.

• Sent: the chat text you are about to share, with emails and phone numbers also removed.
• If it fails: you see Besidka's own warnings and a "check before sharing" reminder.

Past-chat retrieval

Models that cannot call tools cannot search your chat history themselves. For those models, and only while Search and reference chats is on, Jev decides whether your message refers to an earlier, separate conversation. If it does, Besidka searches your past chats in its own database and adds up to three short excerpts, and the reply links the chats it used. A phrase such as "last time" or "we discussed" starts the search without asking Jev.

• Sent: your message and your last three turns.
• If it fails: Besidka searches your past chats anyway, because missing a needed search costs more than one quick lookup. Without the key, or with the check off, there is no such search.

Privacy: what leaves Besidka and where it goes

• Only the checks you leave on. Nothing goes to TypeSafe AI until you save a key, and a switched-off check sends nothing.
• Redaction before every call. Never-stored values are removed before any text leaves Besidka. The share check also removes emails and phone numbers.
• No text in logs. Besidka's logs keep counts, scores, statuses and timing for these checks, never the text.
• Who receives it. TypeSafe AI, Inc. hosts its service in the United States and is not certified under the EU–US Data Privacy Framework. Towards Besidka it is an independent controller, like the AI providers whose keys you add: Besidka has no contract with it about your data. Your own terms with TypeSafe AI apply. Its data processing agreement is made with you as its account holder and relies on the EU Standard Contractual Clauses and the UK Addendum. Its privacy policy says it does not train models on your input. It keeps data "as long as necessary", with no fixed period.
• What it can contain. The text can hold whatever your chats, projects and memory hold, including sensitive information. Your consent before your first message covers sending it to the providers whose keys you add, and TypeSafe AI is one of them.

The Privacy Policy lists what each check sends.

Why it is optional and BYOK

• Your key, your account. TypeSafe AI bills its own account, the one you create, at its own price. Besidka adds nothing on top and holds no TypeSafe AI key of its own.
• Saving the key is the opt-in. All seven checks are on once you save it. Each one has its own switch, and deleting the key stops all of them at once.
• Nothing breaks without it. Besidka works fully without a TypeSafe AI key. Every check except past-chat retrieval falls back to exactly what Besidka does without the key.

How well it works: the eval, with caveats

Before shipping, every check was tested against jev-1.14.0 with data sets written for the test, in English, Ukrainian, Polish and, for most checks, Russian. For the two checks that scored perfectly at first, a second, harder set was written and labelled before any of its results were seen.

Results at the shipped thresholds:

• Auto web search (179 messages, 99 need a search): 99.0% precision, 100% recall; 1 needless search out of 80.
• Memory pre-filter (99 chats, 54 with a lasting fact about the user): 0 of 54 wrongly skipped; 44 of the 45 chats without one skipped.
• Prompt-injection screening (104 snippets, 53 injections, 13 security articles that quote attacks): 94.6% precision, 100% recall; 1 of 13 articles flagged.
• Search re-ranking (8 queries, 20 to 30 candidates each): ranking quality (NDCG@5) 0.99, against 0.42 for keyword order.
• Share check (64 chats, 32 with personal data): 100% precision, 100% recall.
• Past-chat retrieval (80 messages, 39 refer to an earlier chat): 97.5% precision, 100% recall.
• Memory checks (48 short texts, many with sensitive data): all 48 classified correctly; the built-in word rules alone got 39.

Read these numbers with their limits:

• Synthetic and small. One person wrote and labelled every item, and each was run once. A 100% means no error on that set, not a measured error rate. With 54 positive examples, a real miss rate of up to about 7% would still fit.
• Tuned on the same data. The thresholds were picked on these sets, which flatters them. They are starting points and will be re-tuned on real use, from scores Besidka logs without any text.
• Not tested yet. Calls from Besidka's servers (the timings were measured from a developer's computer), real chat traffic, personal data hidden inside pasted code, and text written specifically to fool Jev. TypeSafe AI itself says adversarial text can move Jev's answers, and that languages other than English work less well.
• Search re-ranking ships in a slightly different request shape than the one tested, so its cut-off may need re-tuning.

Built-in checks first, Jev second

Jev complements Besidka's built-in checks and never replaces them:

• Memory. The built-in check, fixed word and pattern rules in English, Ukrainian, Polish and Russian, runs on every way memory can be written, with or without a key. It decides what is never stored. Jev can only refuse more.
• Injection. Screening is an extra signal next to the defences that are always on: memory writes are locked after a web search or another tool result, and retrieved text is passed to the model as quoted data, never as instructions.
• Failures. A failed call never blocks you. Each check falls back to what Besidka does without the key.

For what memory is and how to edit it, see How memory works. For the rest of the BYOK setup, see the BYOK guide.