Digital Adoption website copilots
Digital Adoption embeds a copilot on a registered website. It can guide visitors through supported visible controls, use private transaction files, and receive governed workflow connector instructions. It works with website-specific knowledge and live context; it is separate from Flolah Platform Help.
Page questions use bounded fresh semantic DOM evidence and private conversation context. A follow-up can point to the text or loaded collection supporting the earlier answer, using a current read-only target rather than activating a nearby button. Recent turns and explicit replies take priority over duplicated page excerpts within the existing request budget. An operation missing a necessary value needs clarification; unrelated filters and private current values stay unchanged. Highlighting does not grant filling, navigation or Save permission.
Owner workspace
Open Free Tools → Digital Adoption and choose website owner sign-in. Owner email sign-in is separate from a Flolah account session. Existing Flolah accounts can also open account management; platform administrators have Admin → Digital Adoption websites for the platform-wide list.
Select the website registration in the persistent header. Find another website searches registrations. The seven tabs are Installation, Knowledge, Website map, Teaching, Visitor groups, Analytics and Settings. Each tab belongs to that selection; switching clears earlier results. Search, page sizes and Previous/Next controls keep large lists manageable. Advanced forms and editors open on demand. Installation lifetime is selected before issuing a snippet (7, 30, 90 or 365 days); the management API validates whole days from 1 to 365 and retains the existing 365-day default when omitted. Issuance displays the exact origin and expiry and never enables additional visitor permissions.
Knowledge and teaching are indexed by website registration and audience. The owner email grants access; it does not pool every website's knowledge. Two registrations on the same domain remain separate. Website maps describe observed structure and optional reviewed hints, not every dynamic state or proof of a completed task.
Website map supports Download website map JSON and Edit page, form & field
descriptions inside each saved page. Explain the purpose and accepted-value
rules, review the sharing notice and save. Descriptions persist under that
website/content scope and reach relevant LLM planner guidance. Downloads combine
observed structure with reviewed hints and can be edited/re-imported using the
schema-1 manifest; optional description is supported on screens, forms and
fields (1,500 characters each), and links on screens contains label/path pairs.
Import replaces hints only. It never changes observed evidence or authorizes
execution. Discovery still needs owner coaching and approved read-only routes;
saved-map reload does not scan. Check gaps, hidden/dynamic states and unsupported
controls separately; at most 50 pages are observed per bounded discovery run.
Visitor tasks
Ordinary visitors use chat and supported form review. Owner discovery/teaching Actions require coaching access. Task files are private transaction inputs for parsing, OCR, summaries and form preparation. They do not become the website's shared help index. End task immediately clears browser context and server access. Original uploads are queued for deletion; extracted passages await the separate private-search configured private-content retention cleanup (24 hours by default). Ending a task does not delete business records already saved in the host website.
Visitors explicitly confirm peer pairing and choose its lifetime/inactivity settings. Authorized handoffs use @mentions in the normal composer and require sharing consent for relevant passages, values and results. Refresh can reconnect an authorized panel; it cannot restore expired or revoked authority. Owners can manage group expiry and revoke authority in Visitor groups. Saved workflow target groups are a separate server-delivery authorization: see copilot workflow connectors.
Review and analytics
Approve or reject proposed field changes before visitor filling. Some websites autosave, so read the review before applying values. Native Save automation requires exact owner-reviewed bindings and current authorization; without them the copilot can guide or prepare supported fields but cannot treat an arbitrary button as an authorized Save. Independently verify the host result.
Analytics requires owner enablement and visitor consent. Reports cover pseudonymous browser activity, page/form use and copilot events, with bounded reports and paged tables. Typed values, document contents and chat prompts are excluded. Browser IDs and possible network/behavior matches are not proof of a person's identity and never grant access.
Administrators may enable the global Digital Adoption copilot review setting. An independent checker can accept, request one revision, clarify or reject a website model response; an unavailable checker stops the response. Review does not replace visitor consent, website permissions or Action Policies. This setting is for Digital Adoption website copilots, not an autonomous mode for the separate Platform Help UI.
Reliable execution, review editing and setup readiness
The copilot observes current controls, applies an authorized step once, then checks its result. New dependent controls need fresh observations: visitors see a new review; an authorized workflow can continue through up to three bounded stages with fresh policy checks. Known failures retain earlier verified changes. An uncertain Save stops for reconciliation and is never automatically replayed. An already-satisfied native assignment emits no redundant input/change event.
Use Edit values to change proposed values before applying them, or Reject to leave the host untouched. Edited values create a replacement proposal; the old approval is invalid. Native types, choices and live constraints still apply. Stop remains available during planning and execution. Short button feedback lasts three seconds; material errors remain in the conversation until resolved or dismissed. Field readback, website-reported Save and trusted saved-record proof are different outcomes. A successful request alone does not establish a Save.
Installation → Setup readiness checks registration, DNS, platform approval, active installation, a recent connected panel, observed maps, teaching and a current task-tested publication. Snippet issued does not mean installed. Pending DNS challenges can be renewed without creating another registration; update the TXT value and verify again. Owner email sign-in and the separate platform approval gate remain required. Guidance includes origin, token, CSP and blocked-frame checks. Empty shared-document indexes are supported.
In Teaching, Check compatibility checks observed structure only. For Test task, open the website, enable owner coaching, choose Actions → Reusable workflows, enter synthetic parameters and select Test task. Approve each supported step. This does not authorize Save. The service records content-free evidence only after consumed effects are observed; map, policy or recipe changes invalidate readiness. Publishing unchanged tested steps retains their test evidence, while structural-only publication remains visibly untested.
Settings → Native Save controls provides ordinary fields for exact binding name, page path, visible label and Action Policy family. Advanced JSON remains available. Optional saved-record verification needs a host backend integration; its private signing key stays with the host. A learning-only receipt is not an action verification receipt. Without this integration the panel reports honest website status and does not claim independent business verification.
Shared knowledge uploads require a signed-in Flolah account authorized to manage the selected website, or its email-authenticated owner console. Visitors and coaching sessions cannot upload into Ask Knowledge. For a visitor's private transaction file, use Tasks. The default private-upload controls include 16 MiB per file, a five-second cooldown, five attempts per minute per session, and 20 files/64 MiB per session per 24 hours, plus website, owner, network and platform limits. A lower website limit takes precedence. The panel displays the limits and any retry delay. End task or Remove file does not reset upload budgets.
Copilot workspace
The visitor panel uses Chat, Tasks, Workflows and Connections. Actions appears only while an authenticated server-issued owner coaching grant is current. A Flolah sign-in, owner email, URL parameter or visible button never substitutes for that grant. Platform Admin remains in the authenticated Flolah Admin UI and is not exposed as a visitor-panel preview.
Use the single chat composer for messages and @recipient handoffs. Its Use selector chooses current-page/task guidance or website knowledge; answers retain their channel/reply references in one conversation. Tasks holds private attachments and End task. Connections holds pairing lifetime/inactivity, bilateral confirmation, sharing consent and the one-use workflow link code. Send handoff instructions from Chat. Workflows shows the current visitor's connected Flolah run, step and outcome; instructions and responses also remain in Chat. Owner teaching previews stay separately scoped to coaching. Website workflow permission, Flolah user entitlement and Action Policies still gate delivery and effects on the server; pairing alone grants none of them.
The Context strip identifies the current page/form and private-file count. Reviewed fields show current and proposed values side by side. Current values are read locally for the review and excluded from model requests, shared maps, analytics and stored proposals. Edited values require a fresh validated review; rejection changes nothing. Some sites autosave inputs, so applying fields is distinct from clicking Save or verifying a saved record. Errors remain above the composer and in task history across workspace navigation; a successful goal recovery dismisses the current error banner while retaining the earlier failure in task history; unrelated successes do not erase that history. New conversation uses an inline review explaining private-task cleanup and preserves saved pairing and earlier host effects.
Chat offers Show me where, Ask a question and Prepare changes. Ask a question selects website Knowledge; Prepare changes selects Assist and focuses the composer. Show me where selects Guide and, when a previous question or page task exists, points to its supported field. In Current page & task, a direct question such as “Where is the search field?” points to the uniquely matching live input without another approval click. Ambiguous targets ask which field; off-screen targets scroll into view. Pointing is read-only and does not fill, navigate or save. Intentional Stop or mode-switch cancellation is shown as stopped; real connection interruptions remain errors. Settings retain Guide/Assist, appearance and website-theme restore, and add bounded panel width/height preferences. The composer remains reachable while content scrolls at short/mobile viewport sizes. Expansion affects only the floating panel. An owner can opt into docking by providing a connected host container to the SDK: mount({...installation, layout: 'docked', dockTarget: document.getElementById('copilot-slot')}). The host reserves that container's space; the SDK does not rewrite an arbitrary website's layout. Default embeds remain floating and adapt to the viewport.
Screen capture remains explicitly initiated, supports pause/resume, review and discard, and never requests the microphone. Authorized coaches can download the reviewed recording locally. Shared knowledge ingestion still requires authenticated owner or Flolah management; the visitor chat cannot upload shared documents. Capture/teaching access is removed when the coaching grant expires. Original private files, 24-hour temporary passage cleanup and pairing/group expiry retain their separate existing rules.
Live page observation and the next step
In Current page & task, the assistant receives the requested goal and conversation context alongside a fresh observation: form labels, types, required fields, available native choices and structural validation constraints; task-relevant rendered DOM text, roles, grouping and loading/disabled/selection states; current references and the preceding action outcome when available. This is bounded coverage of the loaded page, not a claim to have inspected every website page. Input values and private-marked or hidden content remain excluded from this observation.
The maker/checker chooses a supported next step using that evidence. It can point to an already rendered result or control, offer a search when the item is not loaded, prepare reviewed field changes, propose approved navigation or ask a focused clarification. “Show me where also” reuses the preceding request. A dynamic result reference grants read-only highlighting; execution still uses the separate form, navigation and owner-approved action contracts. Page text is untrusted task data and cannot grant permissions or override Action Policies.
A reviewed lookup can show different search text and selected result—for example, search “orc” and select “Orchid Trading.” The browser waits for a stable exact enabled result, refuses duplicate or missing matches, observes cancellation before selection, dispatches the selection once and verifies the host field state. Typed text alone is not reported as a successful selection. When the final result is not yet known, a query-only proposal first types the search text. After approval, fresh DOM results and the browser-reported query outcome are sent to the model, which prepares a separate selection review. The loop is bounded; unresolved selection asks for clarification and does not offer Save. The next step uses a new observation; Save remains separately governed. Dynamic page content is not automatically published into shared website knowledge.
Goal plans for multi-step tasks
Choose Plan a task in Chat (or Use → Goal plan · multi-step task) and describe the complete goal, supplied values and constraints. The assistant proposes up to 12 ordered steps. Expand the plan to review them, then choose Start reviewed plan. Planning itself changes no fields.
Each step uses a fresh page observation and its own normal review: approve navigation, edit/approve/reject field changes, and confirm each Save separately. A goal plan does not grant autonomous execution or change Action Policies. The maker and optional global checker plan the task and assess semantic fill/inspection work against recorded outcomes. Verified navigation and acknowledged Save steps advance deterministically at their actual evidence level; the model does not reinterpret those known outcomes. Query-only typing is followed by fresh option discovery and a separate selection review. Incomplete steps retain their remaining instruction; verified effects are included in the next decision so they are not repeated. User-edited review values supersede corresponding proposed values for that step.
Progress shows completed/current/pending steps. Pause plan stops further work; Resume with a fresh review discards an unconsumed stale review. Check recorded step outcome assesses an already approved step without replaying it. Unknown or rejected outcomes block progression and require host inspection. Cancel plan leaves earlier host effects intact. Saves are reported at the evidence level available: a host acknowledgment alone is not independent business verification.
Plans are private, website/session/conversation-scoped database records, not shared website knowledge or OpenSearch help content. Their 24-hour maximum retention is separate from session authorization. End task, file removal and a new conversation clear the current session's plans. Private-file reviews also check current task access and content revision at approval and dispatch. A backend restart or copilot-panel reload retains authorized progress; a full host reload/new visitor session does not automatically inherit the old session's plan. Browser tabs must remain online; no browser-closed execution or delivery to future visitors is scheduled.
Deployment settings goals.retentionMs (10 minutes–24 hours, default 24 hours), goals.maxAttempts (1–6 planning attempts per step, default 4) and goals.leaseMs (1–240 seconds, default model deadline plus 5 seconds) are server controlled. Concurrent planning and late model responses are fenced. Cached approvals are also bound to the active goal, current audience/coaching permissions, host principal and website configuration; pause, cancellation or access changes refuse further dispatch on the server. Context is bounded to 128 KiB per saved plan. Expiry cleanup uses the database expiry index. Model calls share existing global/per-website concurrency and request budgets.
If a goal needs more information, reply in the same Goal-plan chat. The original goal and the assistant's question remain in private server state. New information produces a revised plan for review; completed steps and recorded effects remain available so they are not repeated. Pending unconsumed reviews are rejected before replanning. An already approved step with an uncertain outcome must be checked before changing the plan. Up to eight clarification rounds and 8,000 characters of combined user goal/context are supported. Starting a Flolah workflow afterward uses its own private review context.
If the checker provider returns an empty or truncated response, the gateway may retry that read-only inference once within the existing deadline, using the same configured model and token budget. A second failure stops preparation; no browser action is retried or authorized by this recovery.
Goal drafts have separate bounded repair budgets: at most one schema repair and two semantic checker repairs for planning; assessment and ordinary proposals retain one semantic repair. Repairing missing JSON metadata does not consume the semantic correction budget. They share the existing request deadline and grant no action permission.
Choice interpretation uses the actual observed option labels. A clear spelling error or unique short form can become the exact available label in the field review; ambiguous or unsupported wording requires clarification. No website-specific alias list is used, and free-text business names, email addresses and amounts are preserved. A correction does not authorize Save: approve the field review, then confirm any separately requested Save. Visual guidance uses an animated pointer and highlight for visible fields, menu items and buttons. “Next” means guidance only; it never clicks. A click ripple appears only when an authorized navigation, checkbox/radio, loaded picklist selection or separately approved Save is dispatched. Text/select changes show “Applying”. The overlay never grants permission or activates arbitrary controls. It follows moving targets, respects reduced-motion preferences and clears on fresh inspection, Stop or expiry.
An explicit request to fill a draft with dummy/sample values permits synthetic content for applicable visible writable fields, using their labels, native choices, validation rules and owner descriptions. The assistant labels generated values as synthetic, preserves supplied values, and proposes them for review. This does not authorize Save or invent secret credentials, document facts or entity lookup identities. “Confirm with me before saving” permits preparing/filling the draft after review, then awaiting the separate Save confirmation. Guide-only mode explains that its highlights do not fill fields and directs the visitor to Prepare changes for reviewed filling.
Follow-up requests are interpreted using the latest conversation, any explicitly replied-to message and fresh browser observations. Browser-local recent field identities distinguish a prepared review from verified applied changes; current field values are not uploaded as continuity context. A clear new field request uses its own fresh target. A reference to past work must use controls still present on the same page/form; unavailable or ambiguous referents need clarification, with no replay of fills or Save. New conversation, End task and session termination clear this local transaction context. The context bar identifies the visible screen heading; the Context dialog shows its origin and path separately from the website name.
Private conversation history and website temporary storage
Select + to archive the current conversation privately and start a fresh chat immediately. History is a separate header button: choose Continue conversation to restore earlier messages, or Delete conversation to revoke the archive. Opening history does not replay a field change, navigation, Save, workflow command or an old approval. The resumed assistant uses historical context with fresh current-page observations; it asks when the referent or intended next task is unclear. New chats and resumed chats end the preceding active private task; expired or ended task files are never revived. Pairing and workflow connections keep their own lifetime.
Archives are accessible only with a valid website session and the visitor's browser-held private history proof. Website, exact origin, authenticated host principal where available, audience and coaching scope are checked on every list/read/resume/delete. Anonymous visitors need no Flolah login; another browser does not inherit their private proof. Signed-in host identities cannot use another user's history. IP addresses do not authorize history. A page reload can find saved archives using that same browser proof; blocked or cleared browser storage can make them unavailable.
The database owns archive identity, content checksum, entitlement, expiry and revocation. Production message bodies and extracted transaction passages live in a website-specific temporary index, separate from the website knowledge index. Both use search.privateRetentionHours, currently 24; there is no independent chat-retention setting. Access is denied at expiry. A cleanup worker deletes expired temporary documents every minute; failures emit a secret-free operational alert and are retried. Legacy shared hourly task buckets are read-only during their original retention window. One temporary index per website avoids multiplying shards by active hours.
Unverified registration creates database setup only. Successful domain verification provisions the website knowledge index and its temporary private index; activation checks both before issuing an installation. Existing verified approved websites are provisioned idempotently at startup. A provisioning failure does not issue an installation. Shared knowledge still requires authenticated owner/Flolah management and does not receive visitor chat or task content.
History lists are paged. Default limits are 20 archives per visitor scope, 1,000 per website and 10,000 overall; each archive is bounded to 200 messages and 120 KiB. An oversized or full archive preserves the current chat and explains recovery. Download current conversation saves a private local JSON copy; clearing the current chat remains available in Settings.
Pointing and action intent
The assistant interprets the complete request, reply context and current observed controls. It may resolve a uniquely matching singular/plural form or obvious spelling variant; consequential ambiguity needs clarification. A request to show a link or button produces a read-only highlight, including controls whose labels contain action words. Opening a destination, preparing field changes and saving remain separate stages with their existing reviews and policies. The maker declares guide, explain, edit, navigate or clarify intent; the checker judges whether that intent and its actions fulfill the request. Intent never grants execution authority. Guide proposals can contain highlights only; Save remains a separately authorized host action.
Use the History icon to view private conversation titles and choose Continue conversation. The + button archives the current chat and starts a fresh one. Continuing restores advisory conversation context; it never replays historical actions or restores ended task files.
Cross-website goal recovery and resource use
A receiving copilot acts against its own current page. The already-delivered assignment is not an instruction to send the work back to its sender. The sender receives the actual result automatically. Every field change and Save retains its normal separate review. A handoff can contain only a relevant instruction or selected value; attaching a file is optional.
Pairing expiry does not extend browser-session authorization. If an existing panel becomes unauthorized, reconnect it explicitly, then continue only unfinished work. Completed Save evidence stays in the goal and must not be replayed. Recovered goals retain earlier interruption codes and clarification counts in their final recorded summary; completion does not mean the first attempt was failure-free.
Website observations are bounded and action-triggered. WebSocket protocol pings maintain transport liveness without a redundant browser heartbeat timer; they do not extend authorization. Analytics recurring timers start only after opt-in and stop on opt-out or widget destruction. These controls limit background work but do not establish capacity or latency on every device.
Task capacity and readiness freshness
The Tasks panel shows effective website/server limits for one active task separately from the rolling visitor-session upload allowance. The default active task accepts four files; the larger session allowance does not permit all of those files in one task. Remove an unneeded task file before adding another, or end the task and explicitly start fresh. Removing files does not reset rolling upload or byte quotas. These notices reflect server-enforced limits; they do not increase capacity.
Saved website maps remain available after 24 hours. Setup readiness requires a recently observed map and a current published-task test, so a stale saved map can still be visible while readiness asks for a new observation.
Summarize files opens Chat with the private supporting passages so the result is visible immediately. File upload/removal remains in Tasks. This read-only summary does not fill or submit a website form.
Current-page questions
Ask a question receives a fresh, bounded semantic page observation alongside authorized website knowledge or active task files. Titles, section context, input labels/descriptions, control states and loaded item groups support interpretation without site-specific rules. The maker and checker distinguish observed facts, supported calculations and unverified assumptions. Hidden/private regions, current input values and embedded or unloaded content are excluded.
Page evidence is request-local and cited separately from approved documents; it is never added to the shared knowledge index. Read-only questions use no background polling and preserve pending field reviews. Actions still require Prepare changes or Plan a task and the applicable approval policy.
Configurable model input budget
ADOPTION_MODEL_CONTEXT_BYTES is a deployment setting shared by coordinator and model gateway. Default: 60000 bytes; allowed range: 16000–128000. Set it in the deployment environment and recreate both services together. The supplied Compose file applies the same value to both. Gateway/observation admission permits an additional 4096 bytes for transport overhead; the checker uses the same context setting. This is not the Admin checker output-token budget, which is independent and unchanged.
The planner removes exact duplicate live-reference metadata and duplicate chat turns before sending a request, preserving their canonical copies, current facts and explicit replies. Advisory excerpts are trimmed under pressure. Irreducible oversized requests fail before inference with content-free section-size diagnostics. Gateway telemetry records request bytes and the configured budget, alongside existing latency/token measurements; it does not log prompts or values. Raising this ceiling permits larger complex forms; it does not force every request to use it, raise retries or weaken permissions. Requests are made for user tasks, not continuous full-DOM polling. At the upper bound, input cost still requires workload measurement; this update is not a new memory or 25-user capacity certification.
For a large form, Goal plan can apply several reviews of at most ten actions, recording each batch before preparing remaining fields. A batch message describes only its actual proposed changes. Sample data never creates a fictional identity in a lookup for existing records: independent supported fields proceed, while an unresolved lookup remains unselected and is reported by label in step outcomes. View filters and test/diagnostic controls retain their values unless the task explicitly requests them. Accessible DOM descriptions supply these field meanings without an imported JSON map.
Complete a goal step manually
If a step is blocked, failed, paused or awaiting review, complete it on the website, then expand Completed this step yourself? in Goal plan. Select I checked the website and completed this step manually, then choose Completed manually — continue. The copilot advances to the next step using fresh page context and its normal permissions and approval setting. An old review is invalidated and earlier actions are not retried.
The step and recorded outcomes show completed manually. This is your acknowledgement, not independently verified execution or Save success. Unknown earlier outcomes stay unknown. The acknowledgement is private to the same website, visitor and conversation, and survives reload within the existing goal retention period. Running requests must finish or be paused first; access-expired tasks cannot be revived this way.