# Semantic Search

## Objective

Find records by meaning, so the right result appears without the exact wording.

An embedding-backed retrieval path over the app's searchable records, blended with the existing keyword search and filtered by the same permissions.

## Before You Begin

This feature is being added to an application that already exists and already
works. Do not scaffold a new project, and do not assume a blank slate.

Inspect the codebase first and establish:

- The existing application structure and where code of this kind already lives.
- The framework and version in use.
- The existing design system — colours, spacing, typography, and component conventions.
- Existing UI components you can reuse instead of writing new ones.
- The existing database structure, if this feature needs to persist anything.
- The existing authentication and authorization system, if this feature is user-scoped.
- Dependencies already installed, so you don't add a library that duplicates one.
- The existing test setup and conventions.

Only start writing code once you understand the above. If the application
already implements part of this feature, extend it rather than replacing it.

## Implementation Instructions

1. Choose per record type exactly which fields are embedded, and exclude anything that is hidden, internal, or restricted. A field that is not searchable today must not become searchable by way of an embedding.
2. Apply tenant scoping and per-record permission checks to the retrieved set before results are shown, and over-fetch candidates so filtering does not leave a short page.
3. Blend semantic results with the existing keyword search rather than replacing it. Exact identifiers, quoted phrases, and precise names must still return their exact match at the top.
4. Re-embed a record when its embedded fields change, and store the model or index version alongside each vector so a version change can trigger a controlled rebuild rather than silently mixing incompatible vectors.
5. Ranking of an already-retrieved set is owned by AI Search Result Reranking. This feature owns candidate retrieval and the index; do not add a second ordering pass here.

## UI and UX Requirements

Match the application's existing design system exactly. Reuse its components,
spacing, and typography. This feature should look like it was always there.

## Responsive Requirements

Works on mobile, tablet, and desktop. Touch targets are large enough to hit on a
phone, and nothing overflows horizontally at 320px.

## Accessibility Requirements

- Fully keyboard navigable.
- Correct semantic elements and ARIA roles.
- Visible focus states.
- Meets WCAG AA contrast.
- Dynamic changes are announced to screen readers.
- Respects prefers-reduced-motion.

## Edge Cases

- Embedding a field the user cannot see leaks it through result ordering even when the field itself is never displayed. Decide the embedded field list explicitly per record type.
- Permission and tenant filters must run after retrieval and before display. A vector index treated as pre-filtered will eventually return another tenant's record.
- A user searching for an order number or a quoted phrase wants that exact record, not something conceptually adjacent. Exact matches must not be diluted by semantic neighbours.
- Content changes and index-version changes both invalidate vectors. Handle re-embedding in the background with a visible backlog, and keep serving keyword results while a rebuild is in progress.
- One-word, ambiguous, or non-supported-language queries produce poor semantic matches. Detect them and fall back to keyword behaviour rather than returning confidently irrelevant results.
- Embedding every record on first deploy is a large, slow, and expensive job. Batch it, make it resumable, and cap its rate so it does not starve the rest of the app.
- If the embedding provider is down, search must still work on keywords alone, with no error shown beyond a quiet note that results may be less relevant.

## Testing

Exercise the feature end to end in the running application. Cover every edge case
above, then run the existing test suite and confirm nothing regressed.

## Acceptance Criteria

- [ ] Only the explicitly listed fields per record type are embedded, and restricted fields are excluded.
- [ ] Tenant and permission filters are applied to retrieved candidates before any result is displayed.
- [ ] Exact identifier and quoted-phrase matches still rank first.
- [ ] Content edits and index-version changes queue a re-embed, and vectors record the version that produced them.
- [ ] Short, ambiguous, or unsupported-language queries fall back to keyword search.
- [ ] Search returns keyword results with no visible failure when the embedding provider is unavailable.
- [ ] The feature matches the existing design system.
- [ ] No existing functionality is broken.

## Adaptation Rules

- Match the existing design system. Do not introduce a new colour palette,
  spacing scale, or component library.
- Reuse existing components and utilities wherever they fit.
- Follow the naming, file layout, and code style already present.
- Do not upgrade, replace, or remove existing dependencies to make this
  feature fit. Adapt the feature to the app, not the app to the feature.
- Do not break existing functionality. If a change is genuinely required in
  existing code, make the smallest one that works and say so.
- If something in these instructions conflicts with how the application is
  built, follow the application and explain the deviation.

## Final Verification

Before you report the work as done:

1. Re-read the acceptance criteria above and check each one against what you
   actually built.
2. Run the application and exercise the feature end to end.
3. Run the existing test suite and confirm you have broken nothing.
4. Check the feature on mobile, tablet, and desktop widths.
5. Check keyboard navigation and focus handling.
6. Summarize what changed: files added, files modified, and anything you
   deliberately did differently because of how this application is built.

If any acceptance criterion is unmet, fix it before reporting completion.