@eeacms/volto-eea-chatbot 4.1.3 → 4.1.4

package/AGENTS.md

@@ -84,6 +84,10 @@ and consistency.
 - **Git Workflow:** When committing changes, always use `git status` to review
   changes and then `git add <relevant files changed>` to stage specific files,
   instead of `git add .`.
+- **Documentation & Badges:** Keep a single set of develop badges in `README.md`
+  to avoid badge duplication. Store demo animations in `docs/volto-eea-chatbot.gif`
+  and reference via `https://raw.githubusercontent.com/eea/volto-eea-chatbot/master/docs/volto-eea-chatbot.gif`.
+
 
 For more detailed contribution guidelines, refer to `DEVELOP.md`.
 For release procedures, refer to `RELEASE.md`.

package/CHANGELOG.md

@@ -4,6 +4,29 @@ All notable changes to this project will be documented in this file. Dates are d
 
 Generated by [`auto-changelog`](https://github.com/CookPete/auto-changelog).
 
+### [4.1.4](https://github.com/eea/volto-eea-chatbot/compare/4.1.3...4.1.4) - 12 October 2026
+
+#### :rocket: New Features
+
+- feat(chat): make the evidence window a block setting [Tiberiu Ichim - [`d7bf4aa`](https://github.com/eea/volto-eea-chatbot/commit/d7bf4aa624df6f7dfb933bb7dc5793061a5771b7)]
+- feat(chat): show when a fact-check ran on partial context [Tiberiu Ichim - [`23dddfa`](https://github.com/eea/volto-eea-chatbot/commit/23dddfab0a0e682fff48ba6d84e3df5fd63800fe)]
+- feat(chat): fact-check against real Onyx chunk text, not search blurbs [Tiberiu Ichim - [`e47a593`](https://github.com/eea/volto-eea-chatbot/commit/e47a593b14ba0bac81c28f9a6a9b7d6d77eb8a82)]
+
+#### :bug: Bug Fixes
+
+- fix(claims): put not_enough_info on a lighter green instead of yellow [Tiberiu Ichim - [`52db6bb`](https://github.com/eea/volto-eea-chatbot/commit/52db6bbb45a2509ce57c5a5ad3dca2fa7f135d46)]
+- fix(claims): yellow for not_enough_info, guard for ungenerated colour classes [Tiberiu Ichim - [`f34f613`](https://github.com/eea/volto-eea-chatbot/commit/f34f613b29d41d3769d26d971b33402d82e8c865)]
+- fix(claims): one explanation per answer, verdict colours that match verdicts [Tiberiu Ichim - [`3a017b9`](https://github.com/eea/volto-eea-chatbot/commit/3a017b963c908f1230b23051b096fb1e30c232e2)]
+- fix(claim-modal): replace the score bar with a three-position verdict scale [Tiberiu Ichim - [`6cd595f`](https://github.com/eea/volto-eea-chatbot/commit/6cd595f623cd79003f0291b1f5264640febc0432)]
+- fix(claim-modal): title the modal by the verdict, not the process [Tiberiu Ichim - [`ac9f982`](https://github.com/eea/volto-eea-chatbot/commit/ac9f982e27132c44ce9140e415a2fc06e9e711a2)]
+- fix(proxy): propagate upstream status; treat empty chunk bodies as missing [Tiberiu Ichim - [`e3ccf47`](https://github.com/eea/volto-eea-chatbot/commit/e3ccf47e46805c40b2685dada3f9fe3426bcde24)]
+- fix(chat): render several evidence segments per claim [Tiberiu Ichim - [`ac3d7e1`](https://github.com/eea/volto-eea-chatbot/commit/ac3d7e1d9bde2ab5441731a04aee40ffa9a0e760)]
+
+#### :house: Documentation changes
+
+- docs: specify the i18n pass for the whole add-on [Tiberiu Ichim - [`b37291a`](https://github.com/eea/volto-eea-chatbot/commit/b37291aac16313eb05c092b2cf3bc13998f5a1e2)]
+- docs: remove duplicate badges and fix demo video [Tiberiu Ichim - [`5e29d36`](https://github.com/eea/volto-eea-chatbot/commit/5e29d368bfe17f6da09f0419542e44a8c499aa78)]
+
 ### [4.1.3](https://github.com/eea/volto-eea-chatbot/compare/4.1.2...4.1.3) - 21 September 2026
 
 #### :bug: Bug Fixes
@@ -106,7 +129,6 @@ Generated by [`auto-changelog`](https://github.com/CookPete/auto-changelog).
 #### :house: Internal changes
 
 - chore: bump version to 3.0.0 (EEA design system dependency) [skip ci] [Tiberiu Ichim - [`bc5a8c1`](https://github.com/eea/volto-eea-chatbot/commit/bc5a8c1f72d4e706b096cb96d0d24c2c3f5ac0f8)]
-- chore: [JENKINSFILE] sanitize sonarqube command Refs #305084 [valentinab25 - [`83bf08b`](https://github.com/eea/volto-eea-chatbot/commit/83bf08b7188548a50801b8e464a4d113f0cbbac4)]
 
 #### :house: Documentation changes
 
@@ -184,7 +206,6 @@ Generated by [`auto-changelog`](https://github.com/CookPete/auto-changelog).
 #### :house: Internal changes
 
 - style: Automated code fix [eea-jenkins - [`8a8c3c4`](https://github.com/eea/volto-eea-chatbot/commit/8a8c3c4172a4f669661378cf1b5a3569d85609e6)]
-- chore: [JENKINSFILE] add package version in sonarqube [valentinab25 - [`535d986`](https://github.com/eea/volto-eea-chatbot/commit/535d986b7adc77743a668bc4ac63f835eef58df3)]
 
 #### :hammer_and_wrench: Others
 
@@ -195,7 +216,6 @@ Generated by [`auto-changelog`](https://github.com/CookPete/auto-changelog).
 #### :house: Internal changes
 
 - style: Automated code fix [eea-jenkins - [`fdcd884`](https://github.com/eea/volto-eea-chatbot/commit/fdcd8848fd4c3f990ca5ea021f407487aebd6010)]
-- chore: [JENKINSFILE] use sonarqube branches [EEA Jenkins - [`3d428d7`](https://github.com/eea/volto-eea-chatbot/commit/3d428d72f32c3d05452b0961c76f5db1c416e05c)]
 
 #### :hammer_and_wrench: Others
 

package/README.md

@@ -2,12 +2,6 @@
 
 [![Releases](https://img.shields.io/github/v/release/eea/volto-eea-chatbot)](https://github.com/eea/volto-eea-chatbot/releases)
 
-[![Pipeline](https://ci.eionet.europa.eu/buildStatus/icon?job=volto-addons%2Fvolto-eea-chatbot%2Fmaster&subject=master)](https://ci.eionet.europa.eu/view/Github/job/volto-addons/job/volto-eea-chatbot/job/master/display/redirect)
-[![Lines of Code](https://sonarqube.eea.europa.eu/api/project_badges/measure?project=volto-eea-chatbot&metric=ncloc)](https://sonarqube.eea.europa.eu/dashboard?id=volto-eea-chatbot)
-[![Coverage](https://sonarqube.eea.europa.eu/api/project_badges/measure?project=volto-eea-chatbot&metric=coverage)](https://sonarqube.eea.europa.eu/dashboard?id=volto-eea-chatbot)
-[![Bugs](https://sonarqube.eea.europa.eu/api/project_badges/measure?project=volto-eea-chatbot&metric=bugs)](https://sonarqube.eea.europa.eu/dashboard?id=volto-eea-chatbot)
-[![Duplicated Lines (%)](https://sonarqube.eea.europa.eu/api/project_badges/measure?project=volto-eea-chatbot&metric=duplicated_lines_density)](https://sonarqube.eea.europa.eu/dashboard?id=volto-eea-chatbot)
-
 [![Pipeline](https://ci.eionet.europa.eu/buildStatus/icon?job=volto-addons%2Fvolto-eea-chatbot%2Fdevelop&subject=develop)](https://ci.eionet.europa.eu/view/Github/job/volto-addons/job/volto-eea-chatbot/job/develop/display/redirect)
 [![Lines of Code](https://sonarqube.eea.europa.eu/api/project_badges/measure?project=volto-eea-chatbot&branch=develop&metric=ncloc)](https://sonarqube.eea.europa.eu/dashboard?id=volto-eea-chatbot&branch=develop)
 [![Coverage](https://sonarqube.eea.europa.eu/api/project_badges/measure?project=volto-eea-chatbot&branch=develop&metric=coverage)](https://sonarqube.eea.europa.eu/dashboard?id=volto-eea-chatbot&branch=develop)
@@ -22,7 +16,7 @@
 
 ## Features
 
-https://github.com/user-attachments/assets/f9b5f813-672f-4e4d-81d0-bf2aec35b587
+![Volto Chatbot](https://raw.githubusercontent.com/eea/volto-eea-chatbot/master/docs/volto-eea-chatbot.gif)
 
 The **Volto Chatbot** block allows the integration of an AI-powered chatbot into your Volto project. It offers a customizable interface and advanced settings to tailor the chatbot's behavior to your needs. Below is an overview of its features and configuration options.
 
@@ -334,11 +328,4 @@ See [LICENSE.md](https://github.com/eea/volto-eea-chatbot/blob/master/LICENSE.md
 ## Funding
 
 [European Environment Agency (EU)](http://eea.europa.eu)
-de is European Environment Agency (EEA).
-All Rights Reserved.
-
-See [LICENSE.md](https://github.com/eea/volto-addon-template/blob/master/LICENSE.md) for details.
 
-## Funding
-
-[European Environment Agency (EU)](http://eea.europa.eu)

package/artifacts/I18N-SPEC.md

@@ -0,0 +1,257 @@
+# i18n specification — volto-eea-chatbot
+
+**Status: not implemented.** Written 2026-10-08 for a dedicated branch that handles
+internationalisation across the whole add-on. Nothing here is urgent: no locale is
+filled today, so every string already falls back to English.
+
+Line numbers are from `develop` at the time of writing. Re-run §4's command to
+regenerate the inventory before starting.
+
+---
+
+## 1. Current state (measured)
+
+The Volto add-on i18n stack is present and wired:
+
+| piece | location | state |
+| --- | --- | --- |
+| locale catalogues | `locales/{de,en,it,ro}/LC_MESSAGES/volto.po` | present, **empty** |
+| template | `locales/volto.pot` | present, dated **2026-09-17** (stale) |
+| extractor | `make i18n` → `node_modules/.bin/i18n --addon` | works, **manual** — no CI step |
+| message definitions | `src/AISearchInput/messages.js` (`defineMessages`) | pattern to copy |
+| consumption | `<FormattedMessage {...messages.x} />` | 48 call sites |
+
+Coverage is the problem:
+
+| measure | value |
+| --- | --- |
+| msgids in `volto.pot` | **11** — all from `AISearchInput` |
+| msgstr filled, any locale | **0 of 10** in de, it, ro and en |
+| intl call sites in `ChatBlock/` | **8** (ChatWindow 3, AIMessage 2, AutoResizeTextarea 2) |
+| hardcoded English literals in `ChatBlock/` | **~40** (28 JSX/prop literals + ~12 in JS constants and utils) |
+| fact-check UI strings extracted | **0** |
+
+So the add-on is translatable in principle and English-only in practice. The
+`en/volto.po` being empty is normal and should stay empty — `defaultMessage` is the
+source of truth.
+
+---
+
+## 2. Target design
+
+**One messages file per feature**, matching the existing pattern:
+
+- `src/ChatBlock/messages.js` (new)
+- `src/AISearchInput/messages.js` (exists, keep)
+
+The extractor scans `src/**/messages.js`; keep that exact filename.
+
+**Message ids are the English text** (Volto convention, already used in the `.pot`).
+Do not invent key names for ids — `defaultMessage` carries the copy, `id` is the key.
+
+**Components** use `<FormattedMessage {...messages.x} />`, or `useIntl().formatMessage(...)`
+when the string must be a plain value (e.g. an `aria-label` prop, a `title` attribute).
+
+**Pure helpers return message descriptors, not strings.** This is the decision that
+matters for the fact-check UI: `ChatBlock/utils/index.tsx` maps a verdict to copy and
+is unit-tested without any React or intl context. Returning descriptors keeps it that way.
+
+```js
+// src/ChatBlock/messages.js
+import { defineMessages } from 'react-intl';
+
+export default defineMessages({
+  verdictSupportedHeader: {
+    id: 'Supported by the sources',
+    defaultMessage: 'Supported by the sources',
+  },
+  // …one entry per verdict
+});
+```
+
+```js
+// ChatBlock/utils/index.tsx — descriptor maps instead of string maps
+import messages from '@eeacms/volto-eea-chatbot/ChatBlock/messages';
+
+const VERDICT_HEADER = {
+  supported: messages.verdictSupportedHeader,
+  not_enough_info: messages.verdictNotConfirmedHeader,
+  contradicted: messages.verdictContradictedHeader,
+};
+
+export function claimHeaderLabel(score, contextLimited = false) {
+  if (contextLimited) return messages.verdictUnverifiedHeader;
+  return VERDICT_HEADER[scoreToVerdict(score)];
+}
+```
+
+```jsx
+// ChatBlock/components/markdown/ClaimModal.jsx
+<span className="claim-label">
+  <FormattedMessage {...claimHeaderLabel(claim.score, claim.context_limited)} />
+</span>
+```
+
+Tests then assert on the descriptor (`id` / `defaultMessage`) rather than on English
+prose, so they stay intl-free.
+
+**Plurals and counts move into ICU**, replacing JS string building. Today
+`HalloumiFeedback` composes the banner sentence from a ternary and a template literal;
+that cannot be translated into languages with more than two plural forms.
+
+```js
+// messages.js
+partialContext: {
+  id: '{count, plural, one {# claim was} other {# claims were}} checked against short excerpts, so “not confirmed” may mean missing text rather than a mistake.',
+  defaultMessage:
+    '{count, plural, one {# claim was} other {# claims were}} checked against short excerpts, so “not confirmed” may mean missing text rather than a mistake.',
+},
+```
+
+```jsx
+<MessageContent>
+  <strong><FormattedMessage {...messages.partialContextLead} /></strong>{' '}
+  <FormattedMessage {...messages.partialContext} values={{ count: limitedClaims.length }} />
+</MessageContent>
+```
+
+---
+
+## 3. What must not be extracted
+
+- Answer content, source titles, chunk text — data, not UI copy.
+- `${score}%`, `num_tokens`, and similar interpolations — data.
+- The `claim-quote` block in `ClaimModal`, which is rendered through
+  `dangerouslySetInnerHTML` from the answer itself.
+- Log messages and `debug()` labels.
+- Anything in `services/` or `hooks/` that is only compared or thrown internally —
+  but see `VERIFY_CLAIM_MESSAGES` in `HalloumiFeedback.jsx:12-15`, which **is** shown
+  to the user and must be extracted.
+
+---
+
+## 4. Inventory
+
+Regenerate the candidate list with:
+
+```bash
+cd src && rg -n ">[A-Z“][^<>{}]{3,80}<|(placeholder|title|aria-label|alt|label)=\"?" \
+  ChatBlock AISearchInput -g '!tests' | less
+```
+
+### A. Fact-check UI — 14 strings, added or touched 2026-10-08, zero extracted
+
+| file | line | string |
+| --- | --- | --- |
+| `ChatBlock/utils/index.tsx` | 82–84 | `Supported by the sources` / `Not confirmed by the sources` / `Contradicted by the sources` |
+| `ChatBlock/utils/index.tsx` | 88–90 | `Supported` / `Not confirmed` / `Contradicted` |
+| `ChatBlock/utils/index.tsx` | 99 | `Couldn't be checked` |
+| `ChatBlock/utils/index.tsx` | 104 | `Unverified` |
+| `ChatBlock/components/markdown/ClaimModal.jsx` | 79 | `Fact check` |
+| `ChatBlock/components/markdown/ClaimModal.jsx` | 96 | `Rationale` |
+| `ChatBlock/components/HalloumiFeedback.jsx` | 136 | `Some sources could not be loaded in full.` |
+| `ChatBlock/components/HalloumiFeedback.jsx` | 137–140 | banner sentence with claim count (→ ICU plural) |
+| `ChatBlock/components/HalloumiFeedback.jsx` | 141 | `Every claim was still matched against the text that was available, so this did not change the result.` |
+| `ChatBlock/components/markdown/ClaimSegments.jsx` | 139 | `Jump to Citation` |
+
+### B. Fact-check UI — pre-existing
+
+| file | line | string |
+| --- | --- | --- |
+| `ChatBlock/components/HalloumiFeedback.jsx` | 12–15 | `VERIFY_CLAIM_MESSAGES` (3 progress lines) |
+| `ChatBlock/components/HalloumiFeedback.jsx` | 82 | `Please allow a few minutes for claim verification…` |
+| `ChatBlock/components/HalloumiFeedback.jsx` | 89, 107 | `Fact-check AI answer`, `Retry Fact-check AI answer` |
+| `ChatBlock/components/markdown/Citation.jsx` | 39 | `This doc doesn't have a link.` |
+| `ChatBlock/chat/AIMessage.tsx` | 470, 700 | `See all sources`, `Sources` |
+
+### C. Chat shell
+
+| file | line | string |
+| --- | --- | --- |
+| `ChatBlock/ChatBlockView.jsx` | 45 | `Chatbot` |
+| `ChatBlock/chat/ChatWindow.tsx` | 409 | `Deep research on` (partially done — 3 call sites exist) |
+| `ChatBlock/chat/ChatMessage.tsx` | 26 | `Error` |
+| `ChatBlock/components/FeedbackModal.jsx` | 116 | `Cancel` |
+| `ChatBlock/components/RelatedQuestions.jsx` | 29 | `Related questions:` |
+| `ChatBlock/packets/MultiToolRenderer.tsx` | 267 | `Done` |
+| `ChatBlock/packets/renderers/FetchToolRenderer.tsx` | 35 | `Fetching Documents` |
+| `ChatBlock/packets/renderers/ImageToolRenderer.tsx` | 38 | `Generated Images` |
+
+### D. Accessibility strings — same priority as visible copy
+
+`title` / `aria-label` / `alt` in `ChatMessageFeedback.jsx:39-40`,
+`UserActionsToolbar.jsx:25-26`, `QualityCheckToggle.jsx:15`, `Source.jsx:23`,
+`WebResultIcon.tsx:31`, `AIMessage.tsx:537`, `ChatWindow.tsx:223`. An untranslated
+`aria-label` is a screen-reader defect, not a cosmetic gap.
+
+---
+
+## 5. Testing
+
+- **Existing tests assert English prose.** `utils.test.jsx` (verdict labels),
+  `ClaimModal.test.jsx` (header, badge, `data-verdict`), `HalloumiFeedback.test.jsx`
+  (banner copy, pluralisation). With descriptors, assert `id`/`defaultMessage`
+  instead. Do not wrap these tests in `IntlProvider` — that is the point of returning
+  descriptors.
+- **Component tests and snapshots** will churn: `ClaimModal.test.jsx.snap` and the
+  HalloumiFeedback assertions contain the current English. Expect to re-run with `-u`.
+- **Add a drift guard**: one test that every descriptor exported from
+  `ChatBlock/messages.js` has a matching `id` in `locales/volto.pot`. Cheap, and it is
+  the only thing that stops copy drifting away from the catalogue once translators
+  are involved.
+- Frontend suite command:
+  `CI=true RAZZLE_JEST_CONFIG=src/addons/volto-eea-chatbot/jest-addon.config.js yarn test src/addons/volto-eea-chatbot/src --watchAll=false`
+
+---
+
+## 6. Workflow and CI
+
+`make i18n` is manual and the `.pot` predates the fact-check work, so the catalogue is
+already out of sync with the code. Add a CI step:
+
+```bash
+make i18n
+git diff --exit-code locales/   # fails when a string was added without extracting
+```
+
+Translation fill (de/it/ro) is a separate, human task — the EEA translation workflow,
+not a code task. Until it happens, English fallback applies and nothing changes for
+users.
+
+---
+
+## 7. Suggested commit order
+
+1. `src/ChatBlock/messages.js` + descriptors for the fact-check UI (§4A) and the
+   components that render them. One reviewable commit, no behaviour change.
+2. Banner pluralisation to ICU.
+3. Pre-existing fact-check copy (§4B) and `ClaimSegments`.
+4. Chat shell (§4C) and accessibility strings (§4D).
+5. CI extraction check + regenerated `locales/`.
+
+Each step keeps the suite green on its own; do not mix 1–4 with the `.pot` regeneration.
+
+---
+
+## 8. Gotchas
+
+- **Do not key translations by English text at render time** (no
+  `translate(labelMap[score])` lookup tables). Descriptors only.
+- `utils/index.tsx` is imported by tests with no `IntlProvider` in scope. Descriptors
+  preserve that; passing `intl` into the helpers would force a provider into every test.
+- The extractor only sees `messages.js` files and literal `<FormattedMessage>` /
+  `defineMessages` usage — computed or concatenated strings are invisible to it.
+- Curly quotes in the banner (`“not confirmed”`) must survive extraction; the `.po`
+  files are `charset=utf-8`, so this is fine, but do not let a formatter normalise them.
+- `en/volto.po` stays empty on purpose.
+- Volto core already translates generic words (`Cancel`, `Error`); check for an
+  existing id before adding a duplicate.
+
+---
+
+## 9. Out of scope
+
+- Translating answer text, source titles or chunk content — that is Onyx's corpus.
+- Date/number localisation (`${score}%` formatting) — separate concern, and the
+  percentage is verdict-derived, not a measurement.
+- Filling the de/it/ro catalogues.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@eeacms/volto-eea-chatbot",
-  "version": "4.1.3",
+  "version": "4.1.4",
   "description": "@eeacms/volto-eea-chatbot: Volto add-on",
   "main": "src/index.js",
   "author": "European Environment Agency: IDM2 A-Team",

package/src/AISearchInput/index.js

@@ -3,6 +3,25 @@ import AISearchInputView from './AISearchInputView';
 import AISearchInputEdit from './AISearchInputEdit';
 import { AISearchInputSchema } from './schema';
 
+// The block is offered to Managers everywhere, and to any role on content
+// living under a sandbox path (e.g. /en/sandbox), so that non-Managers can
+// add and experiment with it there.
+export const SANDBOX_PATH = '/sandbox';
+
+export const isAISearchInputRestricted = ({ user, properties }) => {
+  // properties['@id'] is the backend API url of the page being edited
+  const id = properties?.['@id'];
+  if (typeof id === 'string' && id.includes(SANDBOX_PATH)) {
+    return false;
+  }
+
+  if (user?.roles) {
+    return !user.roles.find((role) => role === 'Manager');
+  }
+  // backward compatibility for older Volto versions
+  return false;
+};
+
 export default function installAISearchInputBlock(config) {
   config.blocks.blocksConfig.eeaAISearchInput = {
     id: 'eeaAISearchInput',
@@ -11,12 +30,7 @@ export default function installAISearchInputBlock(config) {
     group: 'common',
     view: AISearchInputView,
     edit: AISearchInputEdit,
-    restricted: ({ user }) => {
-      if (user?.roles) {
-        return !user.roles.find((role) => role === 'Manager');
-      }
-      return false;
-    },
+    restricted: isAISearchInputRestricted,
     mostUsed: false,
     blockHasOwnFocusManagement: false,
     sidebarTab: 1,

package/src/ChatBlock/chat/AIMessage.tsx

@@ -11,10 +11,12 @@ import {
 } from 'semantic-ui-react';
 import { serializeNodes } from '@plone/volto-slate/editor/render';
 import {
+  useChunkEvidence,
   useDeepCompareMemoize,
   useQualityMarkers,
   useScrollonStream,
 } from '@eeacms/volto-eea-chatbot/ChatBlock/hooks';
+import { CHUNK_EVIDENCE_WINDOW } from '@eeacms/volto-eea-chatbot/ChatBlock/services/chunkEvidence';
 import {
   MultiToolRenderer,
   RendererComponent,
@@ -90,14 +92,19 @@ function addQualityMarkersPlugin() {
  * The backend accepts {text, title, source_type} dicts so it can
  * include document titles in LLM prompts without polluting the raw
  * text (which would break span-offset matching).
+ *
+ * `kind` tells the fact-checker how much text it actually got:
+ * `chunk` = real Onyx chunk text, `snippet` = a search blurb, which can
+ * never prove a claim false.
  */
-export function buildHalloumiSource(doc: any, text: string) {
+export function buildHalloumiSource(doc: any, text: string, kind = 'snippet') {
   const cleanedText = text.replace(/\u00A0/g, ' ');
   return {
     text: cleanedText,
     title: doc.semantic_identifier || null,
     source_type: doc.source_type || null,
     link: doc.link || null,
+    kind,
   };
 }
 
@@ -145,10 +152,14 @@ export function getContextSources(
           halloumiSource: buildHalloumiSource(doc, text),
         };
       })
-    : (message.toolCalls || []).reduce(
+    : // `messageProcessor.getMessage()` emits `toolCall` (singular); older
+      // Onyx v2 message shapes used `toolCalls`. Accept both — reading only the
+      // plural made 'all' mode send zero sources, which the fact-checker then
+      // reported as "Answer cannot be verified due to empty sources."
+      [message.toolCall, ...(message.toolCalls || [])].reduce(
         (acc: any, cur: any) => [
           ...acc,
-          ...(cur.tool_result || []).map((doc: any) => {
+          ...((cur && cur.tool_result) || []).map((doc: any) => {
             const cleanedText = (doc.content || '').replace(/\u00A0/g, ' ');
             return {
               ...doc,
@@ -209,6 +220,7 @@ export function AIMessage({
   persona,
   maxContextSegments,
   batchSize,
+  chunkEvidenceWindow = CHUNK_EVIDENCE_WINDOW,
   isLastMessage,
   className = '',
   chatWindowEndRef,
@@ -293,10 +305,56 @@ export function AIMessage({
   // logic intact.
   const showSourcesTab = showSources && !hideSourcesTab;
 
-  const contextSources = getContextSources(
-    message,
-    sources,
-    qualityCheckContext,
+  const contextSources = useMemo(
+    () => getContextSources(message, sources, qualityCheckContext),
+    [message, sources, qualityCheckContext],
+  );
+
+  const qualityControlRequested =
+    messageDisplayed &&
+    qualityCheck &&
+    qualityCheck !== 'disabled' &&
+    forceHalloumi &&
+    showSources &&
+    message.messageId &&
+    message.messageId > -1 &&
+    (qualityCheck === 'enabled' ||
+      qualityCheckEnabled ||
+      verificationTriggered);
+
+  // The Onyx stream only carries search blurbs (~600 chars per document), while
+  // the answer was written from full chunk text. Resolve the real text for
+  // exactly the documents we are about to send, otherwise the fact-checker
+  // marks supported claims as unsupported. See services/chunkEvidence.ts.
+  const { evidence, isFetchingEvidence, evidenceSettled } = useChunkEvidence(
+    contextSources,
+    { enabled: qualityControlRequested, window: chunkEvidenceWindow },
+  );
+
+  const enrichedSources = useMemo(
+    () =>
+      contextSources.map((src: any) => {
+        const resolved = src.document_id ? evidence[src.document_id] : null;
+        if (!resolved || !resolved.text) {
+          return src;
+        }
+        // Same nbsp cleaning as the blurb path: the frontend joins these
+        // strings to highlight segments, the backend indexes the same text, so
+        // both sides must stay byte-identical.
+        const cleanedText = resolved.text.replace(/\u00A0/g, ' ');
+        return {
+          ...src,
+          text: resolved.text,
+          halloumiContext: cleanedText,
+          halloumiSource: {
+            ...(src.halloumiSource || {}),
+            text: cleanedText,
+            kind: resolved.kind,
+          },
+          evidenceChunks: resolved.chunks,
+        };
+      }),
+    [contextSources, evidence],
   );
 
   // Deduplicate sources by text content. The streaming backend often sends
@@ -306,7 +364,7 @@ export function AIMessage({
   // between frontend (span highlighting) and backend (evidence spans).
   const dedupedSources = useMemo(() => {
     const seen = new Set<string>();
-    return contextSources.filter((src: any) => {
+    return enrichedSources.filter((src: any) => {
       const key = src.halloumiContext || src.halloumiSource?.text || '';
       if (key && !seen.has(key)) {
         seen.add(key);
@@ -314,21 +372,12 @@ export function AIMessage({
       }
       return false;
     });
-  }, [contextSources]);
+  }, [enrichedSources]);
 
   const stableContextSources = useDeepCompareMemoize(dedupedSources);
 
-  const doQualityControl =
-    messageDisplayed &&
-    qualityCheck &&
-    qualityCheck !== 'disabled' &&
-    forceHalloumi &&
-    showSources &&
-    message.messageId &&
-    message.messageId > -1 &&
-    (qualityCheck === 'enabled' ||
-      qualityCheckEnabled ||
-      verificationTriggered);
+  // Don't spend a fact-check call on blurbs while the real text is in flight.
+  const doQualityControl = qualityControlRequested && evidenceSettled;
 
   const { markers, isLoadingHalloumi, retryHalloumi }: any = useQualityMarkers(
     doQualityControl,
@@ -345,7 +394,7 @@ export function AIMessage({
     qualityCheckStages,
   );
 
-  const isFetching = isLoadingHalloumi || isLoading;
+  const isFetching = isLoadingHalloumi || isLoading || isFetchingEvidence;
   const halloumiMessage =
     isMessageVerified || doQualityControl ? scoreStage?.label : '';
 
@@ -536,6 +585,7 @@ export function AIMessage({
           showVerifyClaimsButton={showVerifyClaimsButton}
           retryHalloumi={retryHalloumi}
           emptyClaims={emptyClaims}
+          contextQuality={markers?.context_quality}
         />
       )}
 

package/src/ChatBlock/chat/ChatWindow.tsx

@@ -54,6 +54,7 @@ interface ChatWindowProps {
   onDemandInputToggle?: boolean;
   maxContextSegments?: number;
   batchSize?: number;
+  chunkEvidenceWindow?: number;
   onyxVersion?: '2' | '3';
   isPlaywrightTest?: boolean;
   initialQuery?: string | null;
@@ -97,6 +98,7 @@ function ChatWindow({
     onDemandInputToggle = true,
     maxContextSegments = 0,
     batchSize,
+    chunkEvidenceWindow,
     onyxVersion = '3',
     hideSourcesTab,
     extraRemarkPlugins,
@@ -253,6 +255,7 @@ function ChatWindow({
                     persona={persona.id}
                     maxContextSegments={maxContextSegments}
                     batchSize={batchSize}
+                    chunkEvidenceWindow={chunkEvidenceWindow}
                     isLastMessage={index === messages.length - 1}
                     className={
                       index === messages.length - 1 ? 'most-recent' : ''

package/src/ChatBlock/components/HalloumiFeedback.jsx

@@ -73,8 +73,10 @@ const HalloumiFeedback = ({
   sources,
   retryHalloumi,
   emptyClaims,
+  contextQuality,
 }) => {
   const claims = (markers?.claims || []).filter((claim) => !claim.skipped);
+  const limitedClaims = claims.filter((claim) => claim.context_limited);
   const noClaimsScore = claims[0]?.score === null;
   const messageBySource =
     'Please allow a few minutes for claim verification when many references are involved.';
@@ -123,6 +125,23 @@ const HalloumiFeedback = ({
           </MessageContent>
         </Message>
       )}
+
+      {/* The fact-checker only saw part of the answer's sources, so a "not
+          confirmed" verdict may mean missing text rather than a mistake. Said
+          once, at answer level - the claim modal already says "Couldn't be
+          checked" and needs no repeat of this. */}
+      {contextQuality?.level === 'partial' && (
+        <Message color="yellow" className="context-quality-message">
+          <MessageContent>
+            <strong>Some sources could not be loaded in full.</strong>{' '}
+            {limitedClaims.length > 0
+              ? `${limitedClaims.length} ${
+                  limitedClaims.length === 1 ? 'claim was' : 'claims were'
+                } checked against short excerpts, so “not confirmed” may mean missing text rather than a mistake.`
+              : 'Every claim was still matched against the text that was available, so this did not change the result.'}
+          </MessageContent>
+        </Message>
+      )}
     </>
   );
 };