ACTION_ID: ai_classify NAME: Jev: AI Classify CATEGORY: AI CREDITS: token-based, about 0.002 per row (0 with your own TypeSafe key) Answer typed questions about each row with Jev, TypeSafe's decision model. Each question is yes/no, pick one, or score, and every answer comes with a confidence from 0 to 1. All of an action's questions are answered in one Jev call per row. Jev judges only the text you give it. It does not browse the web and does not write text. Put what it needs in `input` (enrichment output, scraped page text, a job title); a domain alone is not enough. 1. INPUTS input (string, required) Input. The row text Jev reads. Supports {{references}}, resolved per row, mixed with labels: "Title: {{input.job_title}}\nCompany: {{enrich_company_linkedin_profile_1.description}}" A row whose input resolves empty fails with "Input is required for Jev: AI Classify" before any credits are charged. questions (raw_array, required) Questions. A JSON array of one or more questions (its JSON string is accepted too). Every question has: id A UUID, unique per question. Keep it when editing: it is what keeps the output columns, and references to them, stable. name The answer column's name. Must stay unique after snake-casing, including the generated " confidence" and " level". type yes_no | pick_one | score question What Jev answers. Literal text: references here are not filled in, so row data goes in `input`. By type: yes_no Optional `yes` / `no`: what counts as each answer. pick_one `options`: 2 to 255 `{label, description?}`, labels unique. `description` tells Jev what the label means. score `levels`: 2 to 10 level names, LOWEST FIRST. personal_api_key (boolean, optional, default false) Use my API key. true runs the row on the organization's own TypeSafe key (a TypeSafe connection on the Floqer Connections page; an agent cannot create it) and costs 0 Floqer credits. With it on and no key, rows fail with "Your TypeSafe API key is missing. Update it in Connections." 2. OUTPUTS Derived by the server from `questions`; there are none until questions are set. Per question, in order: The answer. yes_no: boolean. pick_one: string, the chosen label exactly. score: number from 0 (lowest level) to levels - 1, can be fractional (2 decimals). level score only (string). The level name nearest the score. confidence number, 0 to 1. For yes_no it is the probability of the chosen answer, so never below 0.5. Public names are snake-cased: a question named "Fit" gives {{ai_classify_1.fit}} and {{ai_classify_1.fit_confidence}}. Add Action with inline questions lists the outputs; Configure Action returns only the action_instance_id, so read them back with Get Action or Get Action Outputs. 3. HOW TO CONFIGURE Start from a template or a draft (Get Action Field Options on `questions`): POST /api/v1/workflows/{workflow_id}/sheets/{sheet_id}/actions/{action_instance_id}/options/questions (body: {}) → templates (Industry Classification, Decision Maker Status). Each option's `extras.needs` lists, as {label, hint}, the row data to put in `input`. POST .../options/questions (body: {"context": {"description": "Classify the industry and rate ICP fit", "input": ""}}) → one set of questions drafted from plain words. Nothing is saved or run. description up to 2,000 characters, input up to 20,000. Both return `value` (a JSON string ready for `inputs.questions`) and `extras.questions` (the same array, with fresh UUIDs). Save either. Configure Action body (also accepted inline on Add Action): { "inputs": { "input": "Title: {{input.job_title}}\nCompany: {{input.company_description}}", "questions": [ { "id": "209c4cbe-fbfc-4376-9c1d-ad101d280929", "name": "Sells software", "type": "yes_no", "question": "Does this company sell software?" }, { "id": "5b0f4a52-6c1e-4d8e-9f3a-2e7d1c9b8a41", "name": "Buying role", "type": "pick_one", "question": "What is this person's role in purchasing decisions?", "options": [ { "label": "Decision maker", "description": "C-level, VP, Head of, Director, Founder" }, { "label": "Influencer", "description": "Manager, Lead, senior IC" }, { "label": "Not involved" } ] } ] } } Sending `questions` replaces the whole list. Omit it to leave the questions and their columns untouched. 4. KEY NOTES - Invalid questions are rejected with 400 before anything is saved (missing UUID, empty name or text, wrong option or level count, duplicate labels, names that collide after snake-casing). - Editing questions keeps column IDs by question `id`, so renames and reorders keep downstream references. Deleting a question drops its columns. Deleting or retyping a referenced column returns a `broken_downstream_reference` warning naming the actions to reconfigure. - Route low-confidence rows: gate the next step with `run_if` or a `filter` on `_confidence` (e.g. less than 0.7 goes to review). In the app this is an If-else. - Each row is a single attempt. A timeout (30 s), rate limit or Jev outage fails the cell with the reason; re-run those rows. - Billing: a row holds 0.03 credits, then settles to its token use (about 0.002 credits for a 500-token input). With `personal_api_key` on, 0. 5. WHERE IT FITS IN A WORKFLOW Pattern: enrich or scrape (enrich_person_linkedin_profile, scrape_web_page_using_firecrawl, ...) -> ai_classify on that text -> filter / run_if on the answer and its confidence -> CRM or sequencer push. 6. WHEN TO USE Use ai_classify to label rows with a fixed set of answers: - Job title or persona, seniority, decision-maker status. - Industry or company type from a company description. - Record matching: are these two records the same company or person? - Still at company: does the latest profile still show this employer? - ICP fit scored on a fixed scale. 7. WHEN NOT TO USE The answer needs web research -> action-detail/llm_web_agents.txt (then classify its result here) You need written text (summaries, emails, extracted free-form fields) -> action-detail/llm_models.txt Math, dates, counts or empty checks -> action-detail/formula.txt, action-detail/filter.txt or `run_if` Last updated: 2026-10-05.