What's new in kovac

A page action can now take the visitor to another page of your site, such as a map or a product page, and on a laptop or desktop the chat opens again there so the conversation carries on.

Tools from a connected tool server can now be made available to one workspace only. Other workspaces never see them, and every call tells the tool server which workspace is calling.

Split a website or S3 source into sections, each matched by address and synced on its own schedule: check the news pages every hour and the rest of the site once a week. A section’s sync only crawls its own pages, and the sync log shows each section’s runs.

Choose how precisely a citation highlights its source on your page: the best-matching part, the whole passage, or the exact words a small model picked as the support. Set it on an agent or workflow, and override it for one embed in the Embed tab. On a laptop or desktop, the chat now stays open when a citation takes the visitor to another page of your site.

Sources now sync on the same schedules as automations: every 15 or 30 minutes, hourly, daily, weekly or monthly, at a time of day in your own time zone, with the next runs shown before you save. A source that rarely changes can be checked weekly or monthly instead of every day.

When the chat widget cites a page from your website, tapping the citation now opens that page and highlights the passage the answer came from, instead of showing a text snippet. It is on for new widgets and can be turned on for existing ones in the widget’s settings.

Workspace members can sign in to the trace viewer with their account and see their own workspace’s traces, read-only.

Agents and workflows can now suggest a few questions to ask next, including some that point to other topics your knowledge covers. Turn it on in the agent editor’s Follow-up questions section or in a workflow’s Settings; the suggestions appear under the latest answer in the console chat and in the website chat widget, where they use the widget’s button colours, and one click or tap sends one. The answer itself is unchanged, and API clients still read the suggestions from the reply.

An embedded assistant can now perform actions on your website, such as applying a filter or opening a record, through code your page provides and only after the visitor allows it on a card that names the action and its details. You declare the actions in the widget’s settings, and the install snippet gives you the handler code to fill in. With page control switched on, the assistant can also read the page and propose a short plan of clicks, typed values and same-origin navigations, which the visitor allows on one card.

You can now set what share of answers the online judges grade, for the whole workspace or for a single agent or workflow, and see an estimate of the judge calls a day it means. The scores now say how many answers they are based on.

When you generate a golden set you can now limit it to chosen documents, folders, a date range or a metadata filter, cap the questions from any one document and tune how strictly similar questions are dropped, including against questions already in the set. Save the dialog’s settings as a named preset to start the next run from, and see retrieval quality broken down by question type, phrasing, language, answer length and one- or two-document questions.

An evaluation run now also scores how much of the reference answer your agent’s answer covers, so a reply that is correct but leaves out the explanation shows up. The score is shown beside the others and does not affect whether a run passes.

Golden-set answers are now written from every passage that acceptably answers a question, in the length, style and depth you choose, and you can rewrite the answers of an existing set in one run or regenerate a single one from the review panel. You can also shape the questions themselves: choose their types, difficulty, phrasing and language, write them as a particular kind of person, and add unanswerable questions and questions that need two documents.

An agent can now list the documents in its knowledge collections, by folder, name or details, and read a whole document to summarise it or answer from it. It sees only the collections it is given and the documents the person asking may see.

An agent you allow can now tag or correct a document’s details in its collections. You approve each change unless the agent’s author turns that off, and the agent can only change what the person using it could change.

Questions in a chat can now collect typed details in a small form, such as a name, email or number, and offer an “Other…” choice with a text box. Agents and workflows can both ask this way, and the card checks what you enter before it is sent.

An agent with page reading can open a link or a search result and read the page, including PDFs. Each address is shown to you to approve first, unless the agent’s author turns that off.

Give an agent web search and it can look up current information beyond your collections, citing each page it used as a link. A free search service is included, and an administrator can switch to a paid search provider for more dependable results.

Agents now know the current date, so questions about recent events, deadlines or how long ago something happened get correct answers.

An agent with web search can now use a different search provider from the rest of the workspace, chosen in the agent editor.

The console now shows which version you are using, in the account menu and at the bottom of Home. When a newer version goes live while you have the console open, it asks you to reload.

A workflow can now send every message through its first step, even while an agent is already answering. A router there can decide, message by message, which agent or scripted flow answers.

A new Classify step answers questions about a message — pick a category, yes or no, or a score — and gives each answer a confidence, so a later step can branch on it. Routing steps can now use a fast decision model that answers in a fraction of a second, with a chat model as the fallback.

Admins can erase a document everywhere it is kept — every version, its passages, files, the answers that cited it, and the conversation history and memories that drew on it — and download a signed report of what was removed. Admins can also find a person across the workspace and redact or erase every mention of them, with a signed report of what was done. Deleting can also erase automatically, per workspace or collection, when you turn that on. Erasing also clears the copies kept for monitoring answer quality.

Retrieval quality is checked automatically after you rebuild a collection and every week, and you can compare any two checks side by side. Editors can also run a check from a collection’s Health tab whenever they like.

Turn on health emails for a collection, or for every collection if you are an admin, and you get a message when its health gets worse and another when it is healthy again, at most once an hour unless it gets worse. Admins can also add signed webhooks to post the same alerts to a chat channel or their own service.

Every collection now has a Health tab showing whether its sources sync, its documents ingest, and its search finds the right passages, each against an objective you can tune per collection or for the whole workspace. The Knowledge list shows which collections need attention.

Assign each knowledge gap to a teammate who can fix it, and filter the list to the gaps assigned to you. Turn on the weekly email to get the largest open gaps and the ones you own, with links.

Questions your collections could not answer are now grouped by topic on Knowledge → Gaps, each with a draft of what a covering document should say. Turn a draft into a document in one step, and the topic is checked again once it is in.

A single question step can now ask a whole list of questions, such as a practice round, giving feedback after each answer. When the last answer is in, the person sees a review card with their score and each answer marked right or wrong, with the correct one beside it. While a workflow works, the chat now shows its steps as they run, like an agent’s reasoning.

An agent in a workflow can now hand the conversation over to a flow of steps, such as a scripted practice round that asks its questions one after another without waiting on a model in between. A question step can also show a note above the question, for example feedback on the previous answer.

Visitors can now pin the website widget to the side of the page, come back to their earlier conversations, and search the conversation they have open. You can also set the widget’s language to English or Macedonian, or let it follow each visitor’s browser.

When an integration’s server refuses a tool call, the agent now gets a clear message saying the server returned an error, rather than one that reads as if the server could not be reached. A workflow’s tool step whose call fails now fails, rather than passing the error on as its result. Forge explains that it does not use connected accounts itself and how to talk to an agent that does. A new docs page, Connecting tools, covers adding an integration, signing in with each person’s or the workspace’s account, approvals and troubleshooting.

An agent can now ask a question with answer buttons — yes or no, one or several options, or a scale — and a workflow can pause on such a question until it is answered. The chat and the website widget show it as a card in English or Macedonian, with an optional note above the question such as feedback on the previous answer, and people can always type their answer instead.

Integrations (MCP servers) can now sign in with OAuth, as each person’s own account or as one account the workspace shares, and people connect them from the chat, their Account page or the workspace’s Connections tab. Their tools ask before running unless the integration marks them read-only, which an operator can change per integration and an agent’s author per tool, and a scheduled run stopped at such a question shows as “Waiting for approval or stopped”.

The cost figures on Home and Monitoring now come from the usage kovac itself records for every model call and embedding, over the period you choose.

Choose how often a chat asks before acting: Ask every time, Approve for me or, in Forge, Full access. The choice is enforced by the server, and every change made under Full access is recorded in the audit log.

Upload an image or choose an icon as the embed widget’s avatar. It appears beside every answer.

The embed widget can show visitors what the assistant is doing while it works, as a single status line or as each search and step, with labels you choose.

Dictate a message and have answers read aloud, in Macedonian or English, in the console and in the embed widget. Dictation writes each phrase as you finish it and stops when you go quiet.

The console’s chats share a new message box, with a + menu, a model picker that lists models by name, and in the playground the choice to run one conversation on another model.

Attach images, PDFs, Office documents and text files to a message, in the console’s chats and, when you allow it, in the embed widget on your website.

Forge can create and change agents, skills, workflows and plugins, put an agent on your website, schedule a workflow and start an evaluation. Each change waits for your review, shown as a diff or, for a workflow, on the canvas.

While a workflow runs, the chat shows which steps are running and how far through a list it is, and the running boxes glow on the canvas.

Organise a collection’s documents in folders and move them between folders. An agent can be limited to some folders of a collection.

In the console, an answer keeps being written if you reload the page or lose your connection, and you see it again when you come back. Stop cancels it.

Forge is the console’s chat. Ask it about your workspace, or talk to any of your agents and workflows from the same conversation. It looks things up rather than guessing, and it sees only what you can see.

A website source can follow links through listing pages without storing the listings themselves, and a page reachable at two addresses is stored once, under the canonical address it names.

A workflow can wait until a named person approves it, rejects it or edits the text it is about to use. Requests wait on the Approvals page, and the person who started a run cannot approve it.

A workflow can list and read a collection’s documents and set their metadata, and run your own Python code in an isolated sandbox.

An ingest step adds the pages or text a workflow found to a collection, with the permissions of the person the workflow runs as and within your workspace’s limits.

Run an agent or a workflow on a schedule, or when another system calls its webhook. New steps repeat over a list or until a condition holds, call tools and web APIs, read RSS and Atom feeds and send email, and the API keys they use are stored encrypted. A Scheduled digest template puts them together.

A workflow’s Runs tab lists every run with who started it and what it cost. Open one to see on the canvas what each step did. A new Scripted RAG chat template searches a collection before its agent answers.

Besides agents, a workflow can run fixed steps (a model call, a condition, a router, a search, setting a value, a reply) for the parts of a process that should go the same way every time.

Mark the tools an agent must ask about, and the chat stops before using one, with Allow, Always allow and Deny.

The console’s menu is grouped under Build, Data, Operate and Settings, and each group folds away.

Start a workflow from the Router or Supervisor template, put it on your website like an agent, and see its runs and its cost per agent in Monitoring. Its answers are judged and evaluated the way an agent’s are.

Draw a workflow as boxes and lines, with each agent’s tools, knowledge and skills shown under it. When you try it, the boxes show which agent is working, how many model calls each made and what they cost.

Put several agents together in a workflow: one hands the conversation to another, or delegates a task and uses the result. Edit a workflow as YAML and try it in the chat beside the editor.

Delete an agent, a skill or a whole plugin from the console. Each delete first shows exactly what it will remove, including an agent’s conversations.

Admins invite people to a workspace by email, and each person joins from the link in their invitation.

Verify your company’s domain with a DNS record, and colleagues who sign up with an email address at it can join your workspace, straight away or once an admin approves.

Sign up and sign in with your Google or GitHub account instead of a password.

Create an account on this website, verify your email and get a workspace of your own on a free trial, with no card required.

A collection’s Feedback tab lists the passages behind answers that people rated down or that the automatic quality checks scored low, with the questions that cited them, so you can edit a passage, exclude repeated text or dismiss it.

Documents in a collection that are near copies of each other are flagged, so you can decide which one to keep. Nothing is merged without you, and a pair you mark as not a duplicate stays dismissed.

People can give an answer a thumbs up or down, in the playground and in the embed widget on your website.

Keep earlier versions of a document, set the dates it is in force, and archive what no longer applies, so answers come from the version that is current.

kovac tries several ways of splitting a collection into passages, measures each against your reviewed test questions, and shows which performs better before you apply it.

When a collection declares metadata fields, a question that names a document type, year or other declared value is narrowed to the matching documents automatically.

Each document’s language is detected, keyword search uses language-aware matching, and a question can find documents written in another language.

The playground’s chat has a new layout, with a context meter that shows how much of the model’s context a conversation is using.

Headers, footers and disclaimers repeated across most of a collection’s documents are detected and suggested for exclusion from search, for you to accept.

Images and charts in documents are described in text, and charts are turned into tables where possible, so their content can be searched and cited.

Tables keep their rows, columns and headers through search and citation, and the document viewer shows the original side by side with the passages kovac extracted from it.

Each collection can use its own reranking model, chosen from the models registered in your workspace, and its relevance threshold can be calibrated against your test questions.

Long documents keep their section structure, so an answer can draw on the whole section around the passage that matched, not just the passage alone.

kovac proposes values for a collection’s metadata fields from each document’s text for a reviewer to accept, and each collection can use its own embedding model.

Every change to knowledge is recorded in an audit log you can export to your security tools, storage and ingestion are capped per workspace, and model and storage usage are measured.

When a connected source reports who may read a document, kovac carries that permission over, and a user only receives answers drawn from documents they are allowed to read.

Organise people into departments and decide which departments can use each collection, agent, skill and model.

Connect more of the places your documents live, or import a whole folder or zip archive with optional metadata files alongside it.

Crawl a website from its sitemap or links on a schedule, including pages that need a browser to render.

Every citation opens the exact passage it came from, and sources such as S3-compatible storage stay in sync on a schedule you choose.

kovac reads dozens of formats, from spreadsheets and presentations to email and code, and can generate test questions from a collection’s own documents for your team to review.

Open any document to see what kovac extracted from it, attach metadata that filters search, and re-index a collection while it keeps answering.

Put one of your agents on your own website with a single script tag, a publishable key and per-visitor rate limits.