feat: stream clinical assistant responses

public/js/clinicalAssistant.js

@@ -118,30 +118,11 @@ import { escapeAttr, escapeHtml, renderAssistantMarkdown, stripSourcesSection }
     setBusy(true, 'Searching indexed resources...');
     var loading = appendLoadingMessage('Searching indexed resources', 'Retrieving and synthesizing cited references...');
 
-    fetch('/api/clinical-assistant/chat', {
+    streamAssistantResponse({
-      method: 'POST',
+      message: text,
-      headers: getAuthHeaders(),
+      history: messages.slice(-8),
-      credentials: 'same-origin',
+      includeContext: !includeContext || includeContext.checked
-      body: JSON.stringify({
+    }, loading)
-        message: text,
-        history: messages.slice(-8),
-        includeContext: !includeContext || includeContext.checked
-      })
-    })
-    .then(function (r) { return r.json().then(function (data) { data._status = r.status; return data; }); })
-    .then(function (data) {
-      setBusy(false, 'Ready');
-      if (!data.success) throw new Error(data.error || ('Request failed (' + data._status + ')'));
-      var answer = data.answer || data.markdown || '';
-      lastAnswer = answer;
-      lastSources = data.sources || data.citations || [];
-      replaceLoadingMessage(loading, answer, lastSources, data.suggestions || []);
-      renderSources(lastSources);
-      if (data.model) {
-        var label = document.getElementById('assistant-model-label');
-        if (label) label.textContent = 'Chat: ' + data.model;
-      }
-    })
     .catch(function (err) {
       setBusy(false, 'Error', true);
       replaceLoadingMessage(loading, 'I could not complete the assistant request. ' + err.message + '\n\nThe indexed search service may be busy or temporarily unavailable. Please try again in a moment.');
@@ -149,6 +130,109 @@ import { escapeAttr, escapeHtml, renderAssistantMarkdown, stripSourcesSection }
     });
   }
 
+  async function streamAssistantResponse(payload, loading) {
+    var response = await fetch('/api/clinical-assistant/chat/stream', {
+      method: 'POST',
+      headers: getAuthHeaders(),
+      credentials: 'same-origin',
+      body: JSON.stringify(payload)
+    });
+    if (!response.ok || !response.body) {
+      var fallback = await response.json().catch(function () { return {}; });
+      throw new Error(fallback.error || ('Request failed (' + response.status + ')'));
+    }
+
+    var partial = '';
+    var streamSources = [];
+    var streamSuggestions = [];
+    var lastRender = 0;
+    var bubble = loading ? loading.querySelector('.assistant-bubble') : null;
+    var decoder = new TextDecoder();
+    var buffer = '';
+    var doneData = null;
+
+    function renderPartial(force) {
+      var now = Date.now();
+      if (!force && now - lastRender < 180) return;
+      lastRender = now;
+      if (!bubble) return;
+      loading.classList.remove('assistant-loading-msg');
+      bubble.classList.remove('assistant-thinking');
+      bubble.innerHTML = partial ? renderMarkdown(partial, streamSources) : '<p class="assistant-muted">Generating answer...</p>';
+      var wrap = document.getElementById('assistant-messages');
+      if (wrap) wrap.scrollTop = wrap.scrollHeight;
+    }
+
+    function handleEvent(type, data) {
+      if (type === 'status') {
+        updateLoadingMessage(loading, data.message || 'Working...');
+        return;
+      }
+      if (type === 'sources') {
+        streamSources = data.sources || [];
+        renderSources(streamSources);
+        return;
+      }
+      if (type === 'token') {
+        partial += data.token || '';
+        renderPartial(false);
+        return;
+      }
+      if (type === 'done') {
+        doneData = data || {};
+        streamSuggestions = doneData.suggestions || [];
+        return;
+      }
+      if (type === 'error') throw new Error(data.error || 'Assistant stream failed');
+    }
+
+    var reader = response.body.getReader();
+    while (true) {
+      var chunk = await reader.read();
+      if (chunk.done) break;
+      buffer += decoder.decode(chunk.value, { stream: true });
+      var parts = buffer.split('\n\n');
+      buffer = parts.pop() || '';
+      parts.forEach(function (part) {
+        var parsed = parseSseEvent(part);
+        if (parsed) handleEvent(parsed.type, parsed.data);
+      });
+    }
+    if (buffer.trim()) {
+      var tail = parseSseEvent(buffer);
+      if (tail) handleEvent(tail.type, tail.data);
+    }
+
+    setBusy(false, 'Ready');
+    var answer = (doneData && (doneData.answer || doneData.markdown)) || partial;
+    lastAnswer = answer || '';
+    lastSources = (doneData && (doneData.sources || doneData.citations)) || streamSources;
+    replaceLoadingMessage(loading, lastAnswer, lastSources, streamSuggestions);
+    renderSources(lastSources);
+    if (doneData && doneData.model) {
+      var label = document.getElementById('assistant-model-label');
+      if (label) label.textContent = 'Chat: ' + doneData.model;
+    }
+  }
+
+  function parseSseEvent(block) {
+    var type = 'message';
+    var data = '';
+    String(block || '').split(/\r?\n/).forEach(function (line) {
+      if (line.indexOf('event:') === 0) type = line.substring(6).trim();
+      if (line.indexOf('data:') === 0) data += line.substring(5).trim();
+    });
+    if (!data) return null;
+    try { return { type: type, data: JSON.parse(data) }; }
+    catch (e) { return null; }
+  }
+
+  function updateLoadingMessage(row, detail) {
+    if (!row) return;
+    var el = row.querySelector('.assistant-thinking-detail');
+    if (el) el.textContent = detail;
+  }
+
   function appendLoadingMessage(title, detail) {
     var wrap = document.getElementById('assistant-messages');
     if (!wrap) return null;

src/routes/clinicalAssistant.js

@@ -13,7 +13,7 @@ var { S3Client, GetObjectCommand } = require('@aws-sdk/client-s3');
 var router = express.Router();
 var db = require('../db/database');
 var { authMiddleware } = require('../middleware/auth');
-var { callAI } = require('../utils/ai');
+var { callAI, callAIStream } = require('../utils/ai');
 var { gatewayUrl } = require('../utils/errors');
 var logger = require('../utils/logger');
 var cryptoUtil = require('../utils/crypto');
@@ -249,93 +249,11 @@ router.delete('/clinical-assistant/chats/:id', async function(req, res) {
 router.post('/clinical-assistant/chat', async function(req, res) {
   var started = Date.now();
   try {
-    var message = String(req.body.message || '').trim();
+    var prepared = await prepareAssistantChat(req.body);
-    if (!message) return res.status(400).json({ error: 'Question is required' });
+    if (prepared.direct) return res.json(prepared.direct);
-    if (message.length > 4000) return res.status(400).json({ error: 'Question too long' });
 
-    var chatModel = await getSetting('clinical_assistant.chat_model', '') || await getSetting('models.default', '');
+    var ai = await callAI(prepared.messages, {
-    var searchLimit = clampInt(await getSetting('clinical_assistant.search_limit', '8'), 3, 20, 8);
+      model: prepared.chatModel || undefined,
-    var contextChars = clampInt(await getSetting('clinical_assistant.context_chars', '1400'), 300, 4000, 1400);
-    var behavior = await getSetting('clinical_assistant.system_behavior', DEFAULT_BEHAVIOR) || DEFAULT_BEHAVIOR;
-    var includeContext = req.body.includeContext !== false;
-
-    if (GREETING_RE.test(message)) {
-      return res.json({
-        success: true,
-        answer: 'What clinical question would you like me to look up in the indexed pediatric resources?',
-        sources: [],
-        model: null,
-        search: { skipped: true, reason: 'low_information_input' }
-      });
-    }
-
-    var history = Array.isArray(req.body.history) ? req.body.history.slice(-8) : [];
-    var contextualHistory = priorClinicalHistory(history, message);
-
-    var suggestions = getTopicSuggestions(message);
-    if (!contextualHistory.length && suggestions.length > 0 && message.split(/\s+/).length < 4) {
-      return res.json({
-        success: true,
-        answer: buildSuggestionAnswer(message, suggestions),
-        suggestions: suggestions,
-        sources: [],
-        model: null,
-        search: { skipped: true, reason: 'topic_suggestions' }
-      });
-    }
-
-    var searchQuery = await rewriteSearchQuery(message, history, chatModel).catch(function(e) {
-      console.warn('[clinical-assistant] query rewrite skipped:', e.message);
-      return message;
-    });
-
-    var searchResponse = await semanticSearch(searchQuery, {
-      limit: searchLimit,
-      includeContext: includeContext,
-      contextChars: contextChars
-    });
-    var visualQuery = isVisualSourceQuery(message) || isVisualSourceQuery(searchQuery);
-    var multimodalResponse = visualQuery ? await multimodalSearch(buildMultimodalSearchQuery(searchQuery), { limit: 8 }).catch(function(e) {
-      console.warn('[clinical-assistant] multimodal search skipped:', e.message);
-      return null;
-    }) : null;
-    var rawTextResults = normalizeMcpSearchResponse(searchResponse);
-    var rawMultimodalResults = await classifyAndRerankMultimodalResults(
-      message + ' ' + searchQuery,
-      normalizeMcpMultimodalResponse(multimodalResponse)
-    );
-    var rawResults = rawTextResults.concat(rawMultimodalResults);
-    console.info('[clinical-assistant] retrieval counts:', {
-      text: rawTextResults.length,
-      multimodal: rawMultimodalResults.length,
-      query: searchQuery
-    });
-    var visualSlots = rawMultimodalResults.length ? Math.min(2, Math.max(1, Math.floor(searchLimit / 4))) : 0;
-    var textSlots = searchLimit - visualSlots;
-    var sources = dedupeSources(
-      dedupeSources(rawTextResults).slice(0, textSlots).concat(
-        dedupeSources(rawMultimodalResults).slice(0, visualSlots)
-      )
-    ).slice(0, searchLimit);
-
-    if (sources.length === 0) {
-      return res.json({
-        success: true,
-        answer: 'I could not find relevant indexed Nextcloud resources for that question. Try a more specific clinical term, diagnosis, medication, age group, or textbook topic.',
-        sources: [],
-        model: null,
-        search: { totalFound: 0 }
-      });
-    }
-
-    var context = formatSourcesForPrompt(sources);
-    var messages = [
-      { role: 'system', content: buildSystemPrompt(behavior) },
-      { role: 'user', content: buildUserPrompt(message, context, history, searchQuery) }
-    ];
-
-    var ai = await callAI(messages, {
-      model: chatModel || undefined,
       temperature: 0.15,
       maxTokens: 1800
     });
@@ -348,18 +266,11 @@ router.post('/clinical-assistant/chat', async function(req, res) {
     res.json({
       success: true,
       answer: answer,
-      sources: sanitizeSourcesForClient(sources),
+      sources: sanitizeSourcesForClient(prepared.sources),
-      model: ai.model || chatModel || null,
+      model: ai.model || prepared.chatModel || null,
       provider: ai.provider || null,
       duration: Date.now() - started,
-      search: {
+      search: prepared.search
-        totalFound: rawResults.length,
-        multimodalFound: rawMultimodalResults.length,
-        query: searchQuery,
-        rewritten: searchQuery !== message,
-        verifiedChunkCount: searchResponse.verified_chunk_count || searchResponse.verifiedChunkCount || 0,
-        droppedDocumentCount: searchResponse.dropped_document_count || searchResponse.droppedDocumentCount || 0
-      }
     });
   } catch (e) {
     console.error('[clinical-assistant]', e.message, e.stack || '');
@@ -367,6 +278,59 @@ router.post('/clinical-assistant/chat', async function(req, res) {
   }
 });
 
+router.post('/clinical-assistant/chat/stream', async function(req, res) {
+  var started = Date.now();
+  var streamOpen = false;
+  function sendEvent(type, data) {
+    if (!streamOpen) return;
+    res.write('event: ' + type + '\n');
+    res.write('data: ' + JSON.stringify(data || {}) + '\n\n');
+  }
+  try {
+    res.setHeader('Content-Type', 'text/event-stream; charset=utf-8');
+    res.setHeader('Cache-Control', 'no-cache, no-transform');
+    res.setHeader('Connection', 'keep-alive');
+    if (typeof res.flushHeaders === 'function') res.flushHeaders();
+    streamOpen = true;
+    sendEvent('status', { message: 'Searching indexed resources...' });
+
+    var prepared = await prepareAssistantChat(req.body);
+    if (prepared.direct) {
+      sendEvent('done', Object.assign({ duration: Date.now() - started }, prepared.direct));
+      return res.end();
+    }
+    sendEvent('sources', { sources: sanitizeSourcesForClient(prepared.sources), search: prepared.search });
+    sendEvent('status', { message: 'Generating answer...' });
+
+    var ai = await callAIStream(prepared.messages, {
+      model: prepared.chatModel || undefined,
+      temperature: 0.15,
+      maxTokens: 1800
+    }, function(delta) {
+      sendEvent('token', { token: delta });
+    });
+    var answer = stripModelSourcesSection(String(ai.content || '').trim());
+    logger.audit(req.user.id, 'clinical_assistant_query', 'Clinical assistant streaming query', req, {
+      category: 'clinical', model: ai.model || prepared.chatModel, duration: Date.now() - started
+    });
+    sendEvent('done', {
+      success: true,
+      answer: answer,
+      sources: sanitizeSourcesForClient(prepared.sources),
+      model: ai.model || prepared.chatModel || null,
+      provider: ai.provider || null,
+      duration: Date.now() - started,
+      search: prepared.search
+    });
+    res.end();
+  } catch (e) {
+    console.error('[clinical-assistant stream]', e.message, e.stack || '');
+    if (!streamOpen) return res.status(500).json({ error: assistantErrorMessage(e) });
+    sendEvent('error', { error: assistantErrorMessage(e) });
+    res.end();
+  }
+});
+
 router.post('/clinical-assistant/image', async function(req, res) {
   try {
     var prompt = String(req.body.prompt || '').trim();
@@ -383,6 +347,113 @@ router.post('/clinical-assistant/image', async function(req, res) {
   }
 });
 
+async function prepareAssistantChat(body) {
+  body = body || {};
+  var message = String(body.message || '').trim();
+  if (!message) {
+    var emptyErr = new Error('Question is required');
+    emptyErr.statusCode = 400;
+    throw emptyErr;
+  }
+  if (message.length > 4000) {
+    var longErr = new Error('Question too long');
+    longErr.statusCode = 400;
+    throw longErr;
+  }
+
+  var chatModel = await getSetting('clinical_assistant.chat_model', '') || await getSetting('models.default', '');
+  var searchLimit = clampInt(await getSetting('clinical_assistant.search_limit', '8'), 3, 20, 8);
+  var contextChars = clampInt(await getSetting('clinical_assistant.context_chars', '1400'), 300, 4000, 1400);
+  var behavior = await getSetting('clinical_assistant.system_behavior', DEFAULT_BEHAVIOR) || DEFAULT_BEHAVIOR;
+  var includeContext = body.includeContext !== false;
+
+  if (GREETING_RE.test(message)) {
+    return { direct: {
+      success: true,
+      answer: 'What clinical question would you like me to look up in the indexed pediatric resources?',
+      sources: [],
+      model: null,
+      search: { skipped: true, reason: 'low_information_input' }
+    } };
+  }
+
+  var history = Array.isArray(body.history) ? body.history.slice(-8) : [];
+  var contextualHistory = priorClinicalHistory(history, message);
+  var suggestions = getTopicSuggestions(message);
+  if (!contextualHistory.length && suggestions.length > 0 && message.split(/\s+/).length < 4) {
+    return { direct: {
+      success: true,
+      answer: buildSuggestionAnswer(message, suggestions),
+      suggestions: suggestions,
+      sources: [],
+      model: null,
+      search: { skipped: true, reason: 'topic_suggestions' }
+    } };
+  }
+
+  var searchQuery = await rewriteSearchQuery(message, history, chatModel).catch(function(e) {
+    console.warn('[clinical-assistant] query rewrite skipped:', e.message);
+    return message;
+  });
+  var searchResponse = await semanticSearch(searchQuery, {
+    limit: searchLimit,
+    includeContext: includeContext,
+    contextChars: contextChars
+  });
+  var visualQuery = isVisualSourceQuery(message) || isVisualSourceQuery(searchQuery);
+  var multimodalResponse = visualQuery ? await multimodalSearch(buildMultimodalSearchQuery(searchQuery), { limit: 8 }).catch(function(e) {
+    console.warn('[clinical-assistant] multimodal search skipped:', e.message);
+    return null;
+  }) : null;
+  var rawTextResults = normalizeMcpSearchResponse(searchResponse);
+  var rawMultimodalResults = await classifyAndRerankMultimodalResults(
+    message + ' ' + searchQuery,
+    normalizeMcpMultimodalResponse(multimodalResponse)
+  );
+  var rawResults = rawTextResults.concat(rawMultimodalResults);
+  console.info('[clinical-assistant] retrieval counts:', {
+    text: rawTextResults.length,
+    multimodal: rawMultimodalResults.length,
+    query: searchQuery
+  });
+  var visualSlots = rawMultimodalResults.length ? Math.min(2, Math.max(1, Math.floor(searchLimit / 4))) : 0;
+  var textSlots = searchLimit - visualSlots;
+  var sources = dedupeSources(
+    dedupeSources(rawTextResults).slice(0, textSlots).concat(
+      dedupeSources(rawMultimodalResults).slice(0, visualSlots)
+    )
+  ).slice(0, searchLimit);
+
+  if (sources.length === 0) {
+    return { direct: {
+      success: true,
+      answer: 'I could not find relevant indexed Nextcloud resources for that question. Try a more specific clinical term, diagnosis, medication, age group, or textbook topic.',
+      sources: [],
+      model: null,
+      search: { totalFound: 0 }
+    } };
+  }
+
+  var context = formatSourcesForPrompt(sources);
+  return {
+    message: message,
+    chatModel: chatModel,
+    sources: sources,
+    messages: [
+      { role: 'system', content: buildSystemPrompt(behavior) },
+      { role: 'user', content: buildUserPrompt(message, context, history, searchQuery) }
+    ],
+    search: {
+      totalFound: rawResults.length,
+      multimodalFound: rawMultimodalResults.length,
+      query: searchQuery,
+      rewritten: searchQuery !== message,
+      verifiedChunkCount: searchResponse.verified_chunk_count || searchResponse.verifiedChunkCount || 0,
+      droppedDocumentCount: searchResponse.dropped_document_count || searchResponse.droppedDocumentCount || 0
+    }
+  };
+}
+
 router.post('/clinical-assistant/export-summary', async function(req, res) {
   try {
     if (Array.isArray(req.body.items) && req.body.items.length) {

src/utils/ai.js

@@ -395,6 +395,68 @@ async function callLiteLLM(messages, model, temperature, maxTokens) {
   };
 }
 
+async function assertModelAllowed(requestedModel, options) {
+  options = options || {};
+  if (!requestedModel || options.skipAllowlistCheck) return;
+  try {
+    var db = require('../db/database');
+    var { getAllowedModelIds } = require('./models');
+    var allowed = await getAllowedModelIds(db);
+    if (allowed && allowed.size > 0 && !allowed.has(requestedModel)) {
+      var err = new Error('Model not permitted');
+      err.code = 'model_not_permitted';
+      err.model = requestedModel;
+      throw err;
+    }
+  } catch (e) {
+    if (e && e.code === 'model_not_permitted') throw e;
+    console.warn('[callAI] Allow-list check skipped:', e && e.message);
+  }
+}
+
+async function callAIStream(messages, options, onToken) {
+  options = options || {};
+  var requestedModel = options.model;
+  var model = requestedModel || DEFAULT_MODEL;
+  var temperature = options.temperature || 0.3;
+  var maxTokens = options.maxTokens || 4000;
+  var startTime = Date.now();
+  await assertModelAllowed(requestedModel, options);
+
+  var client = null;
+  var provider = null;
+  if (activeProvider === 'litellm' && litellmClient) {
+    client = litellmClient;
+    provider = 'litellm';
+  } else if (activeProvider === 'openrouter' && openrouter) {
+    client = openrouter;
+    provider = 'openrouter';
+  } else if (activeProvider === 'azure' && azureClient) {
+    client = azureClient;
+    provider = 'azure';
+    model = process.env.AZURE_DEPLOYMENT_NAME || model;
+  }
+  if (!client) throw new Error('Streaming is only configured for OpenAI-compatible providers');
+
+  var content = '';
+  var stream = await client.chat.completions.create({
+    model: model,
+    messages: messages,
+    temperature: temperature,
+    max_tokens: maxTokens,
+    stream: true
+  });
+  for await (var part of stream) {
+    var delta = part && part.choices && part.choices[0] && part.choices[0].delta ? part.choices[0].delta.content : '';
+    if (!delta) continue;
+    content += delta;
+    if (typeof onToken === 'function') onToken(delta);
+  }
+  var duration = Date.now() - startTime;
+  logger.apiCall(null, provider + '/' + model, { model: model, duration: duration, statusCode: 200 });
+  return { success: true, content: content, model: model, provider: provider, duration: duration };
+}
+
 // ============================================================
 // MAIN CALL AI FUNCTION — Routes to correct provider
 // ============================================================
@@ -411,24 +473,7 @@ async function callAI(messages, options) {
   // the budget on a reasoning model outside the configured roster.
   // Falls back silently to DEFAULT_MODEL when no model was provided.
   // Admin test endpoints pass skipAllowlistCheck to test before adding.
-  if (requestedModel && !options.skipAllowlistCheck) {
+  await assertModelAllowed(requestedModel, options);
-    try {
-      var db = require('../db/database');
-      var { getAllowedModelIds } = require('./models');
-      var allowed = await getAllowedModelIds(db);
-      if (allowed && allowed.size > 0 && !allowed.has(requestedModel)) {
-        var err = new Error('Model not permitted');
-        err.code = 'model_not_permitted';
-        err.model = requestedModel;
-        throw err;
-      }
-    } catch (e) {
-      if (e && e.code === 'model_not_permitted') throw e;
-      // DB lookup failure — fall through, let the provider reject if it
-      // doesn't recognize the model. Log for visibility.
-      console.warn('[callAI] Allow-list check skipped:', e && e.message);
-    }
-  }
 
   try {
     var result;
@@ -632,4 +677,4 @@ async function discoverModels() {
   return discovered;
 }
 
-module.exports = { callAI, whisperClient, activeProvider, discoverModels, vertexClient, litellmClient };
+module.exports = { callAI, callAIStream, whisperClient, activeProvider, discoverModels, vertexClient, litellmClient };