AI Assistant

Availability for trial and prepaid workspacesCopy Anchor Link

Unrestricted assistant chat is unavailable for workspaces enrolled in trial or prepaid processing. Enabling a chat channel or purchasing a pack does not remove that restriction. Use the supported public plugin tools, packet intake and review, and public help instead. The channel behavior below applies where ordinary chat is authorized. See processing allowances.

Self-service availabilityCopy Anchor Link

Self-service processing remains unlaunched. Packet-processing credit does not include the general Assistant: metered workspaces cannot start ordinary chat or finding-explanation turns. Connecting Slack or Teams does not enable those turns. The existing assistant controls below apply when your workspace is eligible for chat.

Opening the AssistantCopy Anchor Link

Click the chat button to open the Assistant as a slide-out panel. The header shows the current utility context and includes controls to change width, minimize, or close the panel.

The embedded assistant loads its chat engine when you first open it or choose Explain this finding. You may briefly see Preparing assistant…. A question submitted during loading waits for the engine; you do not need to submit it again. Closing or minimizing the panel keeps the conversation in the current context and does not cancel a pending question.

Panel ControlsCopy Anchor Link

The assistant header provides three controls:

• Width toggle — Cycle between default, medium, and fullscreen widths
• Minimize — Collapse the panel to a floating action button
• Close — Dismiss the panel entirely

Utility ContextCopy Anchor Link

The assistant automatically syncs with your console utility selection. The current utility is displayed in the header bar. If no utility is selected, it shows "All Utilities" and can provide general interconnection guidance.

When a project is selected, document questions use that application's current attachments, including teammate uploads you can access. These are different from your personal File Library: reviewing a packet does not require copying its files into your personal library or relinking them. If access cannot be verified, the assistant says so instead of claiming the packet has no documents.

Explain a Review FindingCopy Anchor Link

Click Explain this finding on a finding to inspect that exact saved review result. The server checks your current project access and loads the source edition saved with that review. It verifies each quotation against that edition before showing it, with its document title and digitized page.

This is a quotation-first explanation. The model can select relevant passages, but cannot write new quotations, section numbers, equipment definitions, or interpretations. Nearby text is labeled separately: being nearby does not prove that it governs the finding. The assistant explicitly leaves the governing section and project applicability unconfirmed. Ask the utility to confirm these before changing your drawing. It does not re-read the drawing or decide pass or fail.

Common-practice and form-model findings are identified as such when a utility quotation cannot be verified. Indexed sources and published requirements are not necessarily human-approved.

Follow Up or Change TopicsCopy Anchor Link

The selected finding appears above the conversation. Typed follow-ups and retries keep that selection, so asking about its source does not turn into a general requirements list. Use Clear finding before changing topics, or select another finding to replace it. Changing the signed-in user, organization, or project clears the selection. Access is checked again on every turn; a saved conversation does not preserve revoked access.

Finding follow-ups use freshly checked source passages, not earlier assistant claims. Clearing a finding starts a clean model conversation; earlier messages remain stored as history but are not reused as evidence or carried into that new conversation.

When Evidence Is UnavailableCopy Anchor Link

The assistant stops the explanation and states the limitation if it cannot verify the exact saved evidence. This can happen with a document review saved before it recorded a per-finding source pin, a review that lacks a requirement snapshot, a rule derived from a form model whose printed field has not been verified, or a missing or changed source edition. Ambiguous quotations and reading limits can leave only partial evidence. In that case the assistant labels the limit and shows only quotations it could verify. If passage selection fails, it falls back to the saved quotations without adding an interpretation. The assistant does not substitute today's requirement or guess from the finding title. Reopen the finding from its review report or ask your reviewer to confirm its applicability.

Supported UtilitiesCopy Anchor Link

The AI Assistant can use general utility context for 1,329+ EIA-backed utilities across the United States. Select a utility and state for scoped guidance.

Specialized extracted requirement coverage depends on the utility procedure documents that have been processed for that utility, state, and interconnection track.

File Upload in ChatCopy Anchor Link

The web chat composer accepts attachments in its interface:

• Formats: PDF or JSON files
• Limit: Up to 5 files per message
• Size: 10MB maximum per file

The current web assistant submit path does not forward the selected file bytes for analysis. Upload the file in a project's Documents view for document processing and compliance validation; do not rely on a chat attachment as evidence that the assistant read the file.

The web chat request itself is bounded before parsing. A serialized chat request over 256 KiB is rejected, and the downstream turn composer applies its own query and history limits. This keeps large client-supplied transcripts from becoming model input or process memory without changing the normal conversation flow.

Suggested QueriesCopy Anchor Link

When starting a new conversation, the assistant may show suggested queries based on your organization type. Utility users see queue-related prompts; developers see compliance and validation prompts.

Quick Answers vs Tool AnswersCopy Anchor Link

The assistant uses quick answers for stable GridWonk platform facts. It should use tools or extracted data for user-specific and utility-specific questions such as uploaded documents, validation results, and requirement lookups.

Quick answers and documentation sources follow the active organization being viewed. A platform administrator impersonating a customer receives the same public-only assistant content as that customer; ending impersonation restores native administrator access.

Conversation sessions and their persisted history are scoped to the signed-in user and active organization. Switching organizations cannot expose another organization's session, and session metadata is not an authorization input.

Published requirement answers do not repeat a document-role label as though it were approved customer wording. When matching rows are still awaiting wording review, the assistant says so; its available-result count includes only requirements it can actually show in chat.

If a requirements lookup returns no rows, the assistant keeps the reason supplied by the published reader. It distinguishes a scope in which every matching row was rejected by the utility from a utility with no technical publication. An unfinished utility review is not a reason for an empty answer — those rows are returned and reported as awaiting a utility decision.

Explicit document-list or count questions can use the query_user_documents fast path. With a selected project, answers report its accessible current attachments. Packet questions about status or recency use tool-backed answers with recorded processing statuses and upload dates; a partial file list is not evidence of a whole-packet status count or review verdict. Explicit File Library or vault questions use your personal library even while a project is selected. Without a project, or when asking without a utility scope, personal inventory behavior is unchanged. None of these results means the assistant has read chat attachment bytes.

An administrator can disable the assistant's Database Access capability. When it is disabled, database-backed utility lookups are unavailable and the assistant should say it cannot perform that lookup instead of returning unscoped or guessed data. The same control applies on every channel.

Slack Assistant Skill SetsCopy Anchor Link

An organization administrator can enable Assistant skill sets from Features and assign versioned instruction bundles for interconnection research, project coordination, or deficiency review. These bundles guide how the Slack assistant structures an answer. They do not add tools, widen access to organization data, or bypass confirmation rules. Each assignment is organization scoped and audited; disabling the feature removes the bundles from later Slack turns.

Set Up the Slack AssistantCopy Anchor Link

An organization administrator opens Features, enables Slack assistant, and chooses Connect Slack under Chat platform connections. After approving the correct workspace in Slack, return to Features and confirm that the connection is Active.

Each member then opens the GridWonk app in Slack and sends a direct message. If prompted, follow the secure account-linking URL while signed in to the organization connected to that workspace. In a channel, invite the app if required, mention @GridWonk, and keep follow-ups in the same thread. Use /gk context <utility> <state> to set channel defaults. /gk status opens the GridWonk console for validation job monitoring; it does not display jobs inside Slack.

For the complete setup, privacy, command, and troubleshooting guide, see GridWonk for Slack.

Slack Agent ViewCopy Anchor Link

When the organization uses the current Slack product-bot manifest, GridWonk conversations open in Agent View. Slack's native stop control cancels the active turn, status updates show progress, and the adapter assigns an automatic session title. Follow-up context comes from the Agent View transcript; ordinary Slack direct-message history exposes only the user's side in this mode.

Changing an existing Slack app to Agent View cannot be reversed. GridWonk schedules that manifest change in a maintenance window and verifies encrypted installation storage, OAuth scopes, event subscriptions, and stop behavior before enabling an organization.

Deep Research in SlackCopy Anchor Link

When GridWonk enables Deep Research for an approved operator organization, start a run with:

/gk research <question> [--sources=web,gridwonk,organization]

The default selects only current, approved procedure documents that back published packet-review requirements for a supported utility. Set utility and state context first. Paid web search is GridWonk-only, disabled by default, and selected only with --sources=web. Private organization documents are never selected by default; add organization explicitly and run the command only in a one-to-one direct message with the app. Public channels, private channels, group direct messages, and Slack Connect cannot use the private lane. A requested source lane must return usable evidence or the run fails instead of silently omitting that lane. GridWonk records provider usage for cost auditing even when a paid web search returns no usable sources and the run fails.

Use /gk research status [run-id] to inspect progress and open the organization-scoped report. Private-run status and stop controls remain in the one-to-one direct message, and only the requester can open a report that used private sources. Use /gk research stop [run-id] to request cancellation before the next durable workflow step. Pilot limits are one active run per person, five per organization, three starts per person per hour, and 20 sources per report.

Validation ActivityCopy Anchor Link

When tool details are enabled, adjacent assistant steps appear together as expandable lines. Each tool line reads Running, Ran, or Failed to run, followed by the tool name. A completed call appears once, with its input parameters and results together when those details are enabled.

Click to expand the tool details that the interface exposes. Tool input and output are useful audit context, but they are not a complete record of the model's internal analysis.

Sources and Docs VisualsCopy Anchor Link

For docs-backed answers, expand Used N sources below the response to see the numbered source list. The numbers match inline citations in the answer. Linked sources open in a new tab; historical evidence without a document URL stays plain text.

Docs sources use the page titles and links returned by the documentation tools. Repeated references to the same URL appear once. Explicit docs citations are limited to paths authorized for the active organization. Customer and impersonated sessions cannot retrieve or cite internal pages. The same list may also include older extracted source links when the response or tool output includes them.

Some docs-backed answers may also include a compact docs visual card, such as a flow, lane, layer diagram, or approved docs image. The answer text carries the same meaning, so the visual is a reading aid rather than a separate source of truth.

Reasoning DisplayCopy Anchor Link

When reasoning display is enabled, the line reads Thinking… while streaming, then Thought with the measured duration when available. Older turns without a measured duration simply read Thought. Expand the line to read the visible explanation; automatic closing follows the display settings, and you can reopen it. Treat it as a visible explanation or progress artifact, not as a complete transcript of hidden internal reasoning.

Message FeedbackCopy Anchor Link

After receiving a response, hover over the message to copy it or mark it as helpful or unhelpful.

Timeout and Error HandlingCopy Anchor Link

If a response takes longer than expected, the assistant shows progress and then a clear error with retry/dismiss controls when available.

Use Cancel request to stop a pending question, including one waiting for the chat engine to load. If the engine itself cannot load, Try Again retries loading even when the conversation is empty. Dismiss clears that loading error without sending a question. If loading keeps failing, reload the page.

Conversation ContextCopy Anchor Link

The assistant carries the selected utility and state, messages in the current chat partition, and an active project or work context when available. For selected-project model answers, a bounded attachment summary is included only after current access and Database Access settings are checked. This summary contains filenames, roles, recorded processing statuses, and upload dates—not document contents. A selected finding additionally loads its authorized saved requirement and source evidence before an explanation; this does not read the applicant's drawing again. Other document and deficiency data still require authorized tool lookups.

Changing the utility, state, or selected project can move you to a different chat partition. Check the context shown in the header and repeat any detail that is missing rather than assuming every prior conversation follows you.

Source-Backed Requirement AnswersCopy Anchor Link

For utility requirement questions, the assistant should prefer extracted requirement rows over generic platform guidance. Include the utility and state when you ask about site control, application fees, forms, or technical requirements. If the assistant does not have enough context, it may ask you to narrow the utility, state, or interconnection track.

To check an uploaded document's processing or review status, open the project's Documents view. To reuse an eligible File Library file, choose Attach from library there instead of uploading a duplicate.

Uploaded applicant documents and utility procedure documents are different sources. Uploaded documents are evidence you provide; procedure documents define the requirements used for validation.

Requirement answers should keep the requested source in view. If extracted rows do not list a detail such as an easement, right-of-way, minimum lease term, exclusivity rule, shared site agreement, milestone, or timing rule, the assistant should say the detail is not listed rather than answer from general interconnection knowledge.

Channel AvailabilityCopy Anchor Link

GridWonk releases each channel per organization. The web console is on by default. The API, CLI and MCP assistant are off by default during the staged release, and Slack and Teams have their own workspace switches. If a channel is off for your organization, the assistant says so and answers nothing.

Every channel — web chat, Slack, Microsoft Teams, the public API, MCP, and the CLI — uses the shared chatbot orchestration core for routing and canonical answers. The surrounding channel adapter, authentication, enabled tools, context supplied, and presentation can differ:

• Same query classification and intelligent routing
• Same canonical FAQ responses (instant, deterministic)
• Utility/state-aware context when the channel supplies it
• Tool-backed requirements and validation routing subject to that channel's authentication and enabled tools
• Same requirements preflight guidance

This shared core improves consistency, but it does not guarantee identical features or output on every channel.

The saved-finding selection described above is a web-console feature. It does not give Slack, Teams, public API, MCP, or CLI callers access through your browser session.

Approved tool-backed assistant turns use Eve. Operators open Eve one channel at a time; a channel that is not enabled returns a temporary unavailable response instead of replaying the request through the Claude Agent SDK. Deterministic quick answers and direct conversational responses keep their existing lanes, and every enabled channel uses the same answer, citation, and terminal-error contract.

For integration details, see Chatbot Channels. For API access, see Chat Streaming API.

Getting the Best ResultsCopy Anchor Link

1. Be specific — "What are Duke Energy NC lease term requirements?" beats "What do I need?"
2. Provide context — Mention your project type and size when relevant
3. Ask follow-ups — Build on previous answers for deeper detail
4. Upload documents in a project's Documents view — The current web chat attachment control does not send file bytes for analysis
5. Name the source you need — Ask "what does the utility require?" for requirements, or "what have I uploaded?" for your own evidence