Victano: EU tenders and grant calls

Stores what this company can actually deliver, which everything else matches against. Read the company's website and ask them the rest first: Victano never crawls a customer's site. Run get_workflow('build-profile') for what to gather: a vague profile produces vague matches. Fields: name/company_name, description, countries, nuts (service area), cpv_hints (cpv or cpv_divisions ok), keywords (stored in description), languages, past_contracts (stored for the bid, not searched; string or brief-shaped object OK), certifications, turnover_eur (annual_turnover_eur ok), headcount (staff_count/staff ok), exclusions, min_value_eur, max_value_eur. Only `countries`, `nuts` and `cpv_hints` narrow it. The value bounds warn on each result instead of hiding rows, and `exclusions` comes back in `your_exclusions` for you to apply.

Begin authentication. Send the user's work email and your conversation's language ('lt', 'en'…); Victano mails a confirm link and a code. Then call connect_verify(request_id) at once: it waits for the click. No card needed.

Finish signing in: call right after connect_start with no code; it waits for the Confirm click (if it returns waiting, call again; never ask the user). Or pass the code they read you. Returns an api_key (shown once) and, without a connection session, a `session` to pass on every later call.

Begins permanent deletion of the connected account: the email address, every company profile, everything in memory, all API keys, and the whole usage history. THIS TOOL DOES NOT DELETE ANYTHING BY ITSELF. It emails the account holder a confirmation link that lasts 24 hours; the account is destroyed only when they open it and confirm on the page. That is deliberate — whoever is typing to you is not necessarily the account holder. BEFORE CALLING: tell the user in your own words exactly what will be erased, that it CANNOT be undone, and that no copy is kept. Get an explicit yes. Do not call this because a user sounded frustrated. Pass cancel=true to call off a pending deletion instead.

Returns the values that EXIST in a slice, with counts: sources, tender vs grant, works/supplies/services, CPV divisions, value bands, countries, languages, status, and `kind` — whether a row can be bid on at all. USE IT BEFORE NARROWING. Guessing a filter value and getting nothing back is indistinguishable from the corpus being empty, so an agent that guesses narrows again, wrongly, and reports absence. Pairs with research_guide: the guide says the slice is too large to survey, this says what it is made of so you know which cut will actually reduce it. Takes the same filters as search_opportunities, so you can explore inside a narrowing you have already applied. Cheap, and no quota.

Everything Victano holds for this account: the working language, the persona that decides which workflows lead, whether memory is on, the tier and remaining quota, and the full contents of memory with where each entry came from. Call it when the user asks what you know about them, what their limits are, or before changing a setting. Nothing here is hidden from the user: if they ask, read it out.

Full normalised record plus the document inventory. Use before advising on a bid; a search card is not enough. Follow with read_document for eligibility text.

The complete step-by-step playbook for one workflow id, including the decision gates, the EU procurement specifics that decide eligibility, and the pitfalls that lose bids. Follow it rather than summarising it.

The decisions recorded for this account, newest first, each with the title, buyer and deadline of the notice it is about — so reading a week back costs this one call rather than one per row. Use it when a user asks what they have already looked at.

What is new or changed since a time (`since` or `since_days`, default 7 days). Scoped to the stored profile by default; scope='all' ranks across everything in the filters. Use it for the weekly check. Search in the notice's own language; research_guide returns the trade lexicon per country.

The playbook library: how bid professionals actually run this work, from the weekly scan through qualification, eligibility screening, reading a specification, clarification questions, and the proposal skeleton. READ A WORKFLOW BEFORE IMPROVISING A PROCESS. Whole jobs (`jobs`) come first. Pass persona='consultant' to see the multi-client ones. Follow with get_workflow.

Rank open tenders and grant calls against the company profile stored with build_profile. Use it to answer 'what fits us'; each result says why it matched (with a NUTS caveat where the region is uncertain) and `coverage` counts the slice. Profile keywords match best in the notices' own language; research_guide returns the trade lexicon per country.

Paged text for one opportunity. Ask for ONE page first: measured over the whole corpus, 95% of notices are a single page and the median holds under 600 characters, because what we store is the notice, not the tender pack. The response tells you `total_pages` and whether there is `more`; ask for the rest only when it says so. Each page is about 4000 characters, and you are charged for pages RETURNED, not pages requested. Page 1 also returns `qualification` where published: selection and award criteria, exclusion grounds and lots — on Lithuanian skelbimas apie pirkimą notices, in the buyer's language.

Records that this company decided about a tender, so it is not re-qualified from scratch next week. Call it whenever the user reaches a conclusion — including a NO, which is the one worth keeping, because a no is twenty minutes of work that would otherwise be repeated. decision: 'bid', 'no_bid', 'watch' or 'conditional'. reason: one of the structured reasons the tool lists. Free text is refused: pick the closest reason and say the detail to the user.

Stores a preference the user has STATED and will hold next week, from the closed list in get_account.storable_keys. Never free text and never about a person. ASK BEFORE STORING, and never store a one-off. forget=['key'] drops keys; forget=[] with nothing else drops everything.

Read this FIRST in a new conversation, before choosing a search strategy. It is generated from the corpus as it stands and from this account's own slice — how many notices are open in their countries today — and the advice CHANGES with that number. Eleven open tenders and four thousand are not the same problem. Also what the relevance score does and does not mean — not a confidence score. Cheap: one call, no quota. Grants: type='grant', countries=['EUR']. Also returns `lexicon`: the trade terms buyers use per country, so you can search in the notice's own language.

Runs 2-5 phrasings of the SAME question and fuses them into one deduplicated ranking. Use it when the right wording is uncertain, which is most of the time here: buyers write in their own language and their own trade terms, so 'school renovation' and 'refurbishment of educational buildings' reach different notices. Each result carries `angles_matched`. A notice found by EVERY phrasing is a strong match; one found by a single phrasing is often a keyword coincidence and worth checking before you recommend it — a distinction you cannot recover by running the searches separately and concatenating them. Give genuinely different angles, not synonyms of one word: the trade term, the plain description, and what the buyer's own department would call it. COST, re-measured 2026-08-29: a median 0.8 s over six calls against 1.5 s for one search, and roughly the same tokens back. Worth it when the wording is genuinely uncertain, wasteful as a default, and it counts one search against your quota per angle.

Search open public tenders and EU grant calls by subject, ranked by relevance, with filters for country, CPV, status, deadline, value and region. Use it to find notices worth a decision, then get_opportunity or read_document. Cards are compact (`summary_cut` marks a cut summary). Search in the notice's own language; research_guide returns the trade lexicon per country.

Sets the working language ('en' or 'lt'), the persona ('supplier', 'bid-writer' or 'consultant', which decides the workflows offered first), and memory ('on', 'paused' or 'off'). ASK THE USER BEFORE CHANGING ANY OF THESE. They are the user's choices, not yours to infer: a user writing one Lithuanian sentence has not asked you to switch their working language. memory='paused' keeps what is stored but stops using it; memory='off' DELETES it immediately and cannot be undone. Say which one you are about to do before you do it.

Send feedback about Victano to the team that builds it: a bug, a wrong answer, missing data, a missing feature, a workflow idea, or praise. Every item is read by a person. When to offer it: an answer came back empty or wrong; the user corrected you or the result; a workflow was missing a step the user needed; the user repeats a manual step the service could do for them; the user says they need something the service does not do. Offer once, in one sentence; the user's task comes first. Ask first. Show the user what you will send and send it only after they agree. Their own words, and any client, case, company or personal detail, go only with their explicit OK; then set confirmed=true. With authored_by='agent' you may report your own observation of the service (the tool, what you expected, what came back) without asking, as long as it holds none of the user's words or confidential content. Fill kind and what (the task, and what went wrong or what is needed). For a wrong answer add expected and got. Add tool and query for context (the query only with the user's OK), severity, and contact_email only when the user wants a reply. The reply carries a reference id. Tell the user it reached the team; do not promise a reply or a date.

Answers 'what is actually out there?' rather than 'what are the best few?'. Returns ONE COMPACT LINE per notice matching the filters — id, title, buyer, deadline — across the whole filtered set rather than a ranked top ten. A survey row costs about a seventh of a search result, so 150 notices here are cheaper than 25 from search_opportunities. USE THIS FIRST when the user's question is about coverage rather than about the best match: 'what is open in our sector', 'is there anything we are missing', 'how much is out there'. Then search_opportunities or get_opportunity on the specific ones worth a decision. Same filters as search_opportunities. The query is optional: with one, rows come back by relevance; without one, by soonest deadline, which is the order a worklist wants. If the scope is larger than the survey can carry, the response says so explicitly and tells you to narrow — it never silently shows you part of the picture. COST, measured 2026-08-25 so you can budget rather than guess: about 50 tokens a row against 349 for a search result. Cheaper PER ROW, not in total — 150 rows is roughly 7,600 tokens, more than a ten-result search. Ask for the breadth you will actually read.

The connected account: tier, persona, corpus freshness, and every limit — what is left in the current window, and the caps that do not renew. Call it when a user asks about limits, when you are unsure whether you are authenticated, or to confirm a setting took.