Knowledge, skills, notes, and tasks

Each organization has an isolated workspace that gives people and agents shared context without making chat history or an external document system the source of truth.

Author with AI

Select Add knowledge or Create skill to open a persisted, organization-scoped assistant thread. While the tenant-bound chat and its first message are being committed, every launch control is disabled and reports Creating governed chat…. A second click cannot create a duplicate thread. The first message records the complete authoring intent and safety boundary rather than a generic placeholder. The assistant:

1. Reads the matching governed authoring guide.
2. Reads the tenant workspace to find overlaps and conflicts.
3. Asks focused questions about intent, sources, triggers, expected output, permissions, and failure behavior.
4. Drafts readable Markdown and, for a skill, proposes realistic test prompts.
5. Runs skill prompts in an isolated, read-only model simulation and displays individual findings.
6. Shows the exact proposed document in a preview card.
7. Waits for a natural person to select Save draft.

Saving runs through Temporal and Fabric Platform's governed mutation pipeline. The assistant cannot activate a record, expand its own permissions, publish content, enable delivery, or send outreach. Assistant-authored records always enter the workspace as drafts.

On the first real model turn, the assistant proposes a short history title through the same governed chat-persistence adapter. When the operator later supplies the actual topic or workflow, the assistant may replace that launch title with a materially more specific one. Every title is checked against the current organization, user, and chat binding before Lakebase accepts it. After each turn, the chat header and history are refetched so they show the latest AI-derived context and actual response. If title generation does not complete, the full initial request remains as the deterministic fallback; GTM Brain never fabricates a completed response or progress event. A turn that can update history allows one additional bounded tool step; other turns retain the standard eight-step limit.

The chat composer accepts one Markdown, MDX, text, or JSON source file up to 50 KB. GTM Brain hashes the text in the browser, labels it as untrusted user-supplied input, and keeps it attached if sending fails so the operator can retry. The source is not treated as verified knowledge merely because it was uploaded. When sent, its bounded text and hash are persisted with the organization-scoped chat message in Lakebase. It does not enter the Knowledge library or its derived Databricks AI Search index until the assistant proposes an exact draft, a natural person approves Save draft, and the record is later activated through the governed workspace lifecycle.

Author manually

Select Write manually to open the structured editor immediately. Knowledge and skill editors provide:

• Preview and Write modes;
• rendered headings, lists, tables, links, emphasis, block quotes, and code;
• workspace or personal scope plus hierarchical folders;
• source metadata, content hashes, AI provider/model/prompt provenance, tags, and skill policy fields;
• structured skill test cases;
• a document viewer with Viewer, Editor, Copy link, and Download actions;
• a contextual assistant bar that can receive bounded Markdown, text, or JSON source files.

Manual creates, updates, and deletes use the same governed mutation path as assistant proposals.

Navigate the resource workspace

Knowledge and Skills use a file-oriented, three-pane workspace:

• Use the compact + action beside the library title to start guided authoring. The larger empty-state actions remain available when a library has no records.
• Workspace and Personal roots keep organization context separate from the current operator's private drafts.
• Hierarchical folders use the record's governed folder value, and documents appear as Markdown files.
• Selecting a folder shows its description, direct files, and a rendered index.md document when the folder contains one.
• The Skills workspace also exposes read-only System packages for Add knowledge and Create skill. Their skill.md files describe the exact authoring workflow and safety boundaries.
• The assistant composer remains bound to the selected file, folder, or system guide. Its context chip makes that binding visible before a message or source file is submitted.

Search filters titles, descriptions, and folder paths while preserving the matching folder hierarchy. Folder and document URLs are shareable within the authenticated organization.

Storage and retrieval

Knowledge and skills are tenant-scoped operational records in the existing Databricks Lakebase PostgreSQL deployment. A save invokes gtm.upsert_knowledge or gtm.upsert_skill through Temporal and Fabric Platform Host. The canonical GtmKnowledgeChanged or GtmSkillChanged event and its immutable snapshot are retained in the Platform Host ledger and mirrored into the GTM event projection. The workspace view folds those ordered events to the latest non-deleted version; it does not write around the governed mutation boundary.

The library search remains a deterministic title, summary, description, and folder filter. When GTM_CHAT_VECTOR_SEARCH=1, targeted assistant and explicitly granted external-agent reads use Databricks AI Search over gtm_brain_analytics.gtm.knowledge_chunks. Lakebase remains authoritative:

1. gtm.upsert_knowledge or gtm.delete_knowledge commits GtmKnowledgeChanged.
2. The action's declared, idempotent knowledge_index adapter projects only active workspace records.
3. A continuous Delta Sync index embeds the derived chunks.
4. Every query carries a tenant filter.
5. GTM Brain discards any hit whose knowledge ID, lifecycle, scope, tenant, or exact content hash no longer matches the canonical Lakebase projection.

Draft, archived, personal, stale, and deleted revisions are absent from the shared semantic index. Results include the Knowledge ID, exact content hash, chunk identity, score, and source URLs so the assistant can cite what it actually retrieved. Active skill execution still binds the exact full knowledge IDs and content hashes admitted for that run; semantic similarity never widens a skill's authority or silently changes its required evidence.

Files supplied to an authoring chat are bounded source inputs, not knowledge records. They become canonical knowledge only after the assistant proposes an exact draft and a natural person selects Save draft. Binary asset storage and retrieval indexes remain adapters; neither is allowed to become an alternative source of business truth.

Current authoring boundaries

GTM Brain intentionally does not give chat a general shell, filesystem, raw SQL, or unrestricted web browser. The current authoring tools can read tenant workspace and strategy projections, read the system authoring guides, run bounded skill simulations, and propose approval-gated knowledge, skill, ICP, and artifact drafts. Uploaded text sources are size-bounded, hashed in the browser, explicitly delimited as untrusted data, and retain their supplied-source metadata. A URL pasted into chat is not claimed as fetched or verified; governed URL fetching, reusable binary assets, and conflict-aware OpenKnowledge editing remain separate future adapters.

Next: Agents, Artifacts, and Models and gateways.