Knowledge

What knowledge is for

Knowledge sources are documents (URLs, manual entries, integration imports) that FixControl's agents use to ground their answers. The intake agent uses them to classify accurately. The support-reply agent uses them to draft replies. The patch developer agent uses them to understand context.

Tenant data is strictly isolated. A search inside your tenant never sees sources from another tenant.

Source kinds

Suggested drafts (human-approved)

FixControl can propose knowledge instead of you writing it from scratch — but a draft is never published until a human approves it. Drafts arrive as pending suggestions with their origin attached, including:

• Resolved Freshdesk tickets — when a support ticket is resolved, FixControl mines the agent's resolution into a draft (quality-gated: too-short, closing-note-only, and generic resolutions are skipped). See Integrations → Freshdesk.
• Resolved FixControl issues — the same harvesting from in-house resolutions.

Each suggestion is deduped against existing knowledge so you don't drown in near-duplicates. Review them on the suggestions surface and approve to promote a draft into a real, indexed source — or reject it. Customer support content is never auto-published.

How retrieval works

When an agent needs context, FixControl searches your knowledge using a mix of semantic and keyword matching, blends the results, and hands the top matches to the agent. You don't tune the formula — you curate the sources.

Operator surface

The Knowledge page is a paginated DataTable with the standard primitives:

• Server-side pagination — large tenants stay snappy; selection survives page changes.
• Saved views (default: All, Mine, Stale, Failed).
• Column visibility + CSV export of the current filter set.
• Keyboard map — j/k/Enter///x/Esc, plus Cmd/Ctrl+A to select every visible row. See Keyboard & command palette for the full list.
• Bulk delete — select rows across pages, confirm once. The operation is audited.
• Sticky chrome — toolbar and column header pin while you scroll long lists.
• Scrape modal stickiness — the Scrape URL modal doesn't dismiss on an accidental click outside it. You can drag a URL onto it, lose focus, switch tabs to grab the URL, and come back without losing half-typed input. Esc, explicit Cancel, and Save all close it.

Adding a source

1. Click Add source or Scrape URL in the toolbar.
2. Choose URL, Manual entry, PDF, or Import from integration.
3. Give it a title and optional tags.
4. The source enters with processing status while FixControl prepares it for search; indexed once retrieval can use it.

Stale sources

Sources can be flagged Stale when their indexing needs refreshing. Click Re-embed in the toolbar to schedule a refresh. Search still works on stale sources via the keyword fallback, but matching is degraded until the refresh completes.

Confidence on knowledge answers

When FixControl uses retrieval to answer (e.g. drafting a SUPPORT reply), the AI result panel shows:

• Confidence of the answer.
• Sources — which knowledge chunks were used, with relevance scores.
• Uncertainty — what the model said it's unsure about.

Click a source to jump to the underlying knowledge entry.

Feedback loop

Every AI answer that uses knowledge has thumbs-up / thumbs-down buttons in the operator UI. The signal feeds the system-learning store, which re-ranks sources over time. Negative feedback also surfaces under the Failed view so an admin can refine.

Bulk delete

Select rows (x to toggle, Shift+x for a range, Cmd/Ctrl+A for the visible page), then Delete in the sticky bulk toolbar. Bulk delete is audited.

Selection survives the next page click — a bulk delete can span an arbitrarily large filtered set without losing the rows you already ticked.

Common mistakes

• Pasting raw email threads as knowledge. They retrieve loudly and crowd out actual answers. Summarise first.
• Letting stale sources accumulate. Run Re-embed to keep retrieval sharp.
• Treating low confidence as "not enough sources". Sometimes there are plenty; the model is unsure for other reasons. Read the rationale, not just the score.

FAQ

What's the difference between codebase sources and userdoc sources? Userdoc sources (manual, scraped, PDF) are added by you and consumed by the support-reply agent. Codebase sources are produced automatically by Inspect codebase and are intended for the dev agents in the Workspace. You don't manage them by hand.

What does the `reference` status mean? Some sources are stored as pointers only — the content isn't indexed, but agents can still cite them by name.

Can I scope a source to a project or codebase? Yes — set the scope when adding. Retrieval honours it.