Skip to content

Speakr is a personal, self-hosted web application designed for transcribing audio recordings

Python

Track embedding API token usage and cost; surface in admin Vector Store tab; add re-embed-all action

Murtaza Nasirfeaf86287789
diff --git a/src/api/admin.py b/src/api/admin.py
index 1dac4f2f..e7750435 100644
--- a/src/api/admin.py
+++ b/src/api/admin.py
@@ -1067,69 +1067,111 @@ def admin_trigger_auto_process():
@admin_bp.route('/admin/inquire/process-recordings', methods=['POST'])
@login_required
def admin_process_recordings_for_inquire():
- """Process all remaining recordings for inquire mode (chunk and embed them)."""
+ """Process recordings for inquire mode (chunk and embed them).
+
+ Default behaviour processes only recordings that do not yet have chunks.
+ Pass ``force=true`` in the JSON body to re-embed every completed
+ recording regardless of existing chunks. This is the migration path after
+ swapping ``EMBEDDING_MODEL`` or ``EMBEDDING_BASE_URL``: the old vectors
+ are deleted and replaced with vectors from the new configuration.
+
+ Body fields:
+ force (bool, optional, default false): re-embed all recordings
+ batch_size (int, optional, default 10): commit cadence
+ max_recordings (int, optional): cap the number processed in one call
+ """
if not current_user.is_admin:
return jsonify({'error': 'Unauthorized'}), 403
-
+
try:
- # Get optional parameters from request
data = request.json or {}
- batch_size = data.get('batch_size', 10)
+ batch_size = int(data.get('batch_size', 10) or 10)
max_recordings = data.get('max_recordings', None)
-
- # Find recordings that need processing
- completed_recordings = Recording.query.filter_by(status='COMPLETED').all()
- recordings_needing_processing = []
-
- for recording in completed_recordings:
- if recording.transcription: # Has transcription
- chunk_count = TranscriptChunk.query.filter_by(recording_id=recording.id).count()
- if chunk_count == 0: # No chunks yet
- recordings_needing_processing.append(recording)
- if max_recordings and len(recordings_needing_processing) >= max_recordings:
- break
-
+ force = bool(data.get('force', False))
+
+ if force:
+ # Re-embed every completed recording with a transcription.
+ recordings_query = Recording.query.filter(
+ Recording.status == 'COMPLETED',
+ Recording.transcription.isnot(None),
+ Recording.transcription != ''
+ )
+ recordings_needing_processing = recordings_query.all()
+ else:
+ # Only process recordings that have no chunks yet.
+ recordings_with_chunks = db.session.query(TranscriptChunk.recording_id).distinct()
+ recordings_needing_processing = Recording.query.filter(
+ Recording.status == 'COMPLETED',
+ Recording.transcription.isnot(None),
+ Recording.transcription != '',
+ ~Recording.id.in_(recordings_with_chunks)
+ ).all()
+
+ if max_recordings:
+ recordings_needing_processing = recordings_needing_processing[:int(max_recordings)]
+
total_to_process = len(recordings_needing_processing)
-
+
if total_to_process == 0:
return jsonify({
'success': True,
- 'message': 'All recordings are already processed for inquire mode.',
+ 'message': (
+ 'No recordings to re-embed.' if force
+ else 'All recordings are already processed for inquire mode.'
+ ),
'processed': 0,
- 'total': 0
+ 'total': 0,
+ 'force': force,
})
-
- # Process recordings in batches
+
processed = 0
failed = []
-
+
for recording in recordings_needing_processing:
try:
success = process_recording_chunks(recording.id)
if success:
processed += 1
- current_app.logger.info(f"Admin API: Processed chunks for recording: {recording.title} ({recording.id})")
+ current_app.logger.info(
+ f"Admin API: {'Re-embedded' if force else 'Processed chunks for'} "
+ f"recording: {recording.title} ({recording.id})"
+ )
else:
failed.append({'id': recording.id, 'title': recording.title, 'reason': 'Processing returned false'})
except Exception as e:
current_app.logger.error(f"Admin API: Failed to process recording {recording.id}: {e}")
failed.append({'id': recording.id, 'title': recording.title, 'reason': str(e)})
-
- # Commit after each batch
+
if processed % batch_size == 0:
db.session.commit()
-
- # Final commit
+
db.session.commit()
-
+
+ # On a successful force re-embed, refresh the stored identifier so the
+ # mismatch warning clears.
+ if force and processed > 0 and not failed:
+ try:
+ from src.models import SystemSetting
+ from src.services.embeddings import EMBEDDING_IDENTIFIER
+ SystemSetting.set_setting(
+ key='embedding_identifier',
+ value=EMBEDDING_IDENTIFIER,
+ description='Identifier of the embedding provider and model that produced the stored chunk vectors. Used to detect dimensionality and semantic-space mismatches at startup.',
+ setting_type='string',
+ )
+ except Exception as e:
+ current_app.logger.warning(f"Failed to refresh embedding_identifier after re-embed: {e}")
+
+ verb = 'Re-embedded' if force else 'Processed'
return jsonify({
'success': True,
- 'message': f'Processed {processed} out of {total_to_process} recordings.',
+ 'message': f'{verb} {processed} out of {total_to_process} recordings.',
'processed': processed,
'total': total_to_process,
- 'failed': failed
+ 'failed': failed,
+ 'force': force,
})
-
+
except Exception as e:
current_app.logger.error(f"Error in admin process recordings endpoint: {e}")
db.session.rollback()
@@ -1138,48 +1180,84 @@ def admin_process_recordings_for_inquire():
@admin_bp.route('/admin/inquire/status', methods=['GET'])
-@login_required
+@login_required
def admin_inquire_status():
"""Get the status of recordings for inquire mode."""
if not current_user.is_admin:
return jsonify({'error': 'Unauthorized'}), 403
-
+
try:
+ from src.services.embeddings import (
+ EMBEDDING_MODEL, EMBEDDING_BASE_URL, EMBEDDING_IDENTIFIER,
+ USE_API_EMBEDDINGS, EMBEDDING_DIMENSIONS,
+ )
+ from src.models import SystemSetting, TokenUsage
+ from sqlalchemy import func
+
# Count total completed recordings
total_completed = Recording.query.filter_by(status='COMPLETED').count()
-
+
# Count recordings with transcriptions
recordings_with_transcriptions = Recording.query.filter(
Recording.status == 'COMPLETED',
Recording.transcription.isnot(None),
Recording.transcription != ''
).count()
-
+
# Count recordings that have been processed for inquire mode
processed_recordings = db.session.query(Recording.id).join(
TranscriptChunk, Recording.id == TranscriptChunk.recording_id
).distinct().count()
-
- # Count recordings that still need processing
- completed_recordings = Recording.query.filter_by(status='COMPLETED').all()
- need_processing = 0
-
- for recording in completed_recordings:
- if recording.transcription: # Has transcription
- chunk_count = TranscriptChunk.query.filter_by(recording_id=recording.id).count()
- if chunk_count == 0: # No chunks yet
- need_processing += 1
-
- # Get total chunks and embeddings count
+
+ # Count recordings that still need processing (have transcription but no chunks)
+ need_processing = db.session.query(func.count(Recording.id)).filter(
+ Recording.status == 'COMPLETED',
+ Recording.transcription.isnot(None),
+ Recording.transcription != '',
+ ~Recording.id.in_(db.session.query(TranscriptChunk.recording_id).distinct())
+ ).scalar() or 0
+
+ # Total chunks across the system
total_chunks = TranscriptChunk.query.count()
-
+
+ # Detect embedded chunks that pre-date the current configuration. The
+ # stored identifier is what produced the existing vectors; if it does
+ # not match the current EMBEDDING_IDENTIFIER, those chunks are stale
+ # and Inquire mode will return wrong results until they are re-embedded.
+ stored_identifier = SystemSetting.get_setting('embedding_identifier', None)
+ if stored_identifier is None:
+ legacy = SystemSetting.get_setting('embedding_model_name', None)
+ stored_identifier = f"local::{legacy}" if legacy else None
+
+ identifier_mismatch = bool(
+ stored_identifier and stored_identifier != EMBEDDING_IDENTIFIER and total_chunks > 0
+ )
+
+ # Embedding-API usage aggregates (lifetime totals across all users).
+ usage_rows = TokenUsage.query.filter_by(operation_type='embedding').all()
+ embedding_usage = {
+ 'total_tokens': sum(u.total_tokens for u in usage_rows),
+ 'total_cost': sum((u.cost or 0.0) for u in usage_rows),
+ 'request_count': sum(u.request_count for u in usage_rows),
+ }
+
return jsonify({
'total_completed_recordings': total_completed,
'recordings_with_transcriptions': recordings_with_transcriptions,
'processed_for_inquire': processed_recordings,
'need_processing': need_processing,
'total_chunks': total_chunks,
- 'embeddings_available': EMBEDDINGS_AVAILABLE
+ 'embeddings_available': EMBEDDINGS_AVAILABLE,
+ # Embedding configuration surfaced to the UI
+ 'embedding_model': EMBEDDING_MODEL,
+ 'embedding_provider': 'api' if USE_API_EMBEDDINGS else 'local',
+ 'embedding_base_url': EMBEDDING_BASE_URL or None,
+ 'embedding_dimensions_override': EMBEDDING_DIMENSIONS,
+ 'embedding_identifier': EMBEDDING_IDENTIFIER,
+ 'embedding_identifier_stored': stored_identifier,
+ 'embedding_identifier_mismatch': identifier_mismatch,
+ # Embedding-API usage (only meaningful in API mode; local mode = 0)
+ 'embedding_usage': embedding_usage,
})
except Exception as e:
diff --git a/src/services/embeddings.py b/src/services/embeddings.py
index dd427a0b..421c9c35 100644
--- a/src/services/embeddings.py
+++ b/src/services/embeddings.py
@@ -115,8 +115,15 @@ def get_embedding_api_client():
return _embedding_api_client
-def _api_embed(texts):
- """Call the OpenAI-compatible embeddings endpoint and return numpy vectors."""
+def _api_embed(texts, user_id=None):
+ """Call the OpenAI-compatible embeddings endpoint and return numpy vectors.
+
+ When ``user_id`` is provided and the response includes a ``usage`` block,
+ record the call against the daily token-usage aggregate so admins can see
+ embedding cost alongside chat/summarization. Cost is captured when the
+ provider returns it (OpenRouter populates ``usage.cost``); for providers
+ that omit it, only token counts are recorded.
+ """
client = get_embedding_api_client()
if client is None or not texts:
return []
@@ -125,6 +132,29 @@ def _api_embed(texts):
if EMBEDDING_DIMENSIONS is not None:
kwargs['dimensions'] = EMBEDDING_DIMENSIONS
response = client.embeddings.create(**kwargs)
+
+ if user_id is not None and getattr(response, 'usage', None) is not None:
+ try:
+ usage = response.usage
+ # Pydantic v2 stores non-schema fields under model_extra; fall
+ # back to attribute access for SDKs that surface them directly.
+ extras = getattr(usage, 'model_extra', None) or {}
+ prompt_tokens = getattr(usage, 'prompt_tokens', 0) or 0
+ total_tokens = getattr(usage, 'total_tokens', prompt_tokens) or prompt_tokens
+ cost = extras.get('cost') if 'cost' in extras else getattr(usage, 'cost', None)
+ from src.services.token_tracking import token_tracker
+ token_tracker.record_usage(
+ user_id=user_id,
+ operation_type='embedding',
+ prompt_tokens=int(prompt_tokens),
+ completion_tokens=0,
+ total_tokens=int(total_tokens),
+ model_name=EMBEDDING_MODEL,
+ cost=float(cost) if cost is not None else None,
+ )
+ except Exception as track_err:
+ current_app.logger.warning(f"Failed to record embedding usage: {track_err}")
+
return [np.array(d.embedding, dtype=np.float32) for d in response.data]
except Exception as e:
current_app.logger.error(f"Embedding API call failed: {e}")
@@ -181,7 +211,7 @@ def chunk_transcription(transcription, max_chunk_length=500, overlap=50):
-def generate_embeddings(texts):
+def generate_embeddings(texts, user_id=None):
"""
Generate embeddings for a list of texts.
@@ -190,6 +220,9 @@ def generate_embeddings(texts):
Args:
texts (list): List of text strings
+ user_id (int, optional): User to attribute the API call to for token
+ and cost tracking. Local-mode calls do not consume billable usage
+ and ignore this argument.
Returns:
list: List of embedding vectors as numpy arrays, or empty list if no
@@ -199,7 +232,7 @@ def generate_embeddings(texts):
return []
if USE_API_EMBEDDINGS:
- return _api_embed(texts)
+ return _api_embed(texts, user_id=user_id)
if not LOCAL_EMBEDDINGS_AVAILABLE:
current_app.logger.warning("Embeddings not available - skipping embedding generation")
@@ -275,18 +308,18 @@ def process_recording_chunks(recording_id):
recording = db.session.get(Recording, recording_id)
if not recording or not recording.transcription:
return False
-
+
# Delete existing chunks for this recording
TranscriptChunk.query.filter_by(recording_id=recording_id).delete()
-
+
# Create chunks
chunks = chunk_transcription(recording.transcription)
-
+
if not chunks:
return True
-
- # Generate embeddings
- embeddings = generate_embeddings(chunks)
+
+ # Generate embeddings (recording owner gets billed for API-mode usage)
+ embeddings = generate_embeddings(chunks, user_id=recording.user_id)
# Store chunks in database
for i, (chunk_text, embedding) in enumerate(zip(chunks, embeddings)):
@@ -437,9 +470,11 @@ def semantic_search_chunks(user_id, query, filters=None, top_k=5):
current_app.logger.info("scikit-learn not installed - using basic text search as fallback")
return basic_text_search_chunks(user_id, query, filters, top_k)
- # Generate embedding for the query (via API or local model)
+ # Generate embedding for the query (via API or local model). Attribute
+ # the API call to the searching user so embedding cost tracking shows
+ # who issued the query.
if USE_API_EMBEDDINGS:
- api_vectors = _api_embed([query])
+ api_vectors = _api_embed([query], user_id=user_id)
if not api_vectors:
return basic_text_search_chunks(user_id, query, filters, top_k)
query_embedding = api_vectors[0]
diff --git a/templates/admin.html b/templates/admin.html
index 03a4849f..c514cdac 100644
--- a/templates/admin.html
+++ b/templates/admin.html
@@ -857,11 +857,45 @@ Summarization Instructions:
</div>
<div>
<span class="text-[var(--text-muted)]">${t('adminDashboard.embeddingModel')}:</span>
- <span class="ml-2 text-[var(--text-primary)]">all-MiniLM-L6-v2</span>
+ <span class="ml-2 text-[var(--text-primary)] break-all">${ inquireStatus.embedding_model || 'all-MiniLM-L6-v2' }</span>
</div>
<div>
+ <span class="text-[var(--text-muted)]">Provider:</span>
+ <span class="ml-2 text-[var(--text-primary)]">
+ <template v-if="inquireStatus.embedding_provider === 'api'">
+ API (${ inquireStatus.embedding_base_url })
+ </template>
+ <template v-else>
+ Local (sentence-transformers)
+ </template>
+ </span>
+ </div>
+ <div v-if="inquireStatus.embedding_dimensions_override">
<span class="text-[var(--text-muted)]">${t('adminDashboard.vectorDimensions')}:</span>
- <span class="ml-2 text-[var(--text-primary)]">384</span>
+ <span class="ml-2 text-[var(--text-primary)]">${ inquireStatus.embedding_dimensions_override }</span>
+ </div>
+ </div>
+ </div>
+
+ <!-- Identifier mismatch alert: chunks were embedded with a previous configuration -->
+ <div v-if="inquireStatus.embedding_identifier_mismatch" class="mb-6 p-4 rounded-lg bg-[var(--bg-warning-light)] border border-[var(--border-warning)]">
+ <div class="flex items-start">
+ <i class="fas fa-exclamation-triangle text-[var(--text-warning-strong)] mt-0.5 mr-3"></i>
+ <div class="flex-1">
+ <p class="font-semibold text-[var(--text-warning-strong)]">Embedding configuration changed</p>
+ <p class="text-sm text-[var(--text-secondary)] mt-1">
+ ${ inquireStatus.total_chunks } stored chunks were embedded with
+ <code class="px-1 rounded bg-[var(--bg-tertiary)]">${ inquireStatus.embedding_identifier_stored }</code>,
+ but the active configuration is
+ <code class="px-1 rounded bg-[var(--bg-tertiary)]">${ inquireStatus.embedding_identifier }</code>.
+ Inquire mode will return wrong results until you re-embed all recordings.
+ </p>
+ <button @click="reembedAllRecordings"
+ :disabled="isProcessingRecordings"
+ class="mt-3 px-4 py-2 bg-[var(--bg-warning-strong)] text-white rounded-lg shadow hover:opacity-90 disabled:opacity-50 disabled:cursor-not-allowed">
+ <i :class="isProcessingRecordings ? 'fas fa-spinner fa-spin' : 'fas fa-sync-alt'" class="mr-2"></i>
+ ${ isProcessingRecordings ? 'Re-embedding...' : 'Re-embed all recordings now' }
+ </button>
</div>
</div>
</div>
@@ -937,6 +971,29 @@ Summarization Instructions:
<i class="fas fa-tasks text-3xl text-[var(--text-accent)] opacity-50"></i>
</div>
</div>
+
+ <!-- Embedding API usage (only meaningful in API mode) -->
+ <div v-if="inquireStatus.embedding_provider === 'api'" class="bg-[var(--bg-secondary)] p-6 rounded-lg border border-[var(--border-primary)] shadow-sm">
+ <div class="flex items-center justify-between">
+ <div>
+ <p class="text-sm text-[var(--text-muted)]">Embedding API tokens</p>
+ <p class="text-2xl font-semibold text-[var(--text-primary)]">${ (inquireStatus.embedding_usage && inquireStatus.embedding_usage.total_tokens || 0).toLocaleString() }</p>
+ <p class="text-xs text-[var(--text-muted)] mt-1">${ inquireStatus.embedding_usage && inquireStatus.embedding_usage.request_count || 0 } API requests</p>
+ </div>
+ <i class="fas fa-coins text-3xl text-[var(--text-accent)] opacity-50"></i>
+ </div>
+ </div>
+
+ <div v-if="inquireStatus.embedding_provider === 'api'" class="bg-[var(--bg-secondary)] p-6 rounded-lg border border-[var(--border-primary)] shadow-sm">
+ <div class="flex items-center justify-between">
+ <div>
+ <p class="text-sm text-[var(--text-muted)]">Embedding API cost</p>
+ <p class="text-2xl font-semibold text-[var(--text-primary)]">$${ (inquireStatus.embedding_usage && inquireStatus.embedding_usage.total_cost || 0).toFixed(4) }</p>
+ <p class="text-xs text-[var(--text-muted)] mt-1" v-if="inquireStatus.embedding_usage && inquireStatus.embedding_usage.total_cost === 0">Provider did not return cost</p>
+ </div>
+ <i class="fas fa-dollar-sign text-3xl text-[var(--text-accent)] opacity-50"></i>
+ </div>
+ </div>
</div>
<!-- Processing Actions -->
@@ -985,6 +1042,19 @@ Summarization Instructions:
<p class="text-[var(--text-success-strong)] font-medium">${t('adminDashboard.allRecordingsProcessed')}</p>
<p class="text-sm text-[var(--text-muted)] mt-1">${t('adminDashboard.vectorStoreUpToDate')}</p>
</div>
+
+ <!-- Re-embed all (always available) -->
+ <div class="mt-6 pt-4 border-t border-[var(--border-primary)]">
+ <p class="text-sm text-[var(--text-secondary)] mb-3">
+ Re-embed every completed recording with the current configuration. Use this after switching <code class="px-1 rounded bg-[var(--bg-tertiary)]">EMBEDDING_MODEL</code> or <code class="px-1 rounded bg-[var(--bg-tertiary)]">EMBEDDING_BASE_URL</code>. Existing chunks are deleted and replaced.
+ </p>
+ <button @click="reembedAllRecordings"
+ :disabled="isProcessingRecordings || (inquireStatus.processed_for_inquire || 0) === 0"
+ class="px-4 py-2 bg-[var(--bg-secondary)] text-[var(--text-secondary)] border border-[var(--border-secondary)] rounded-lg shadow hover:bg-[var(--bg-tertiary)] focus:outline-none focus:ring-2 focus:ring-[var(--border-focus)] focus:ring-offset-2 transition duration-150 ease-in-out disabled:opacity-50 disabled:cursor-not-allowed">
+ <i :class="isProcessingRecordings ? 'fas fa-spinner fa-spin' : 'fas fa-sync-alt'" class="mr-2"></i>
+ ${ isProcessingRecordings ? 'Re-embedding...' : 'Re-embed all recordings' }
+ </button>
+ </div>
</div>
</div>
@@ -3039,6 +3109,34 @@ Summarization Instructions:
}
};
+ const reembedAllRecordings = async () => {
+ if (isProcessingRecordings.value) return;
+ if (!confirm('Re-embed every completed recording with the current configuration? Existing chunks will be deleted and rebuilt. This may take a while and (in API mode) will incur cost.')) {
+ return;
+ }
+ isProcessingRecordings.value = true;
+ processingResult.value = null;
+ try {
+ const response = await fetch('/admin/inquire/process-recordings', {
+ method: 'POST',
+ headers: {
+ 'Content-Type': 'application/json',
+ 'X-CSRFToken': document.querySelector('meta[name="csrf-token"]').getAttribute('content')
+ },
+ body: JSON.stringify({ force: true })
+ });
+ if (!response.ok) throw new Error('Failed to re-embed recordings');
+ const data = await response.json();
+ processingResult.value = data;
+ await loadInquireStatus();
+ } catch (error) {
+ console.error('Error re-embedding:', error);
+ processingResult.value = { success: false, message: 'Failed to re-embed recordings.' };
+ } finally {
+ isProcessingRecordings.value = false;
+ }
+ };
+
const processBatchRecordings = async (batchSize) => {
if (isProcessingRecordings.value) return;
@@ -3308,6 +3406,7 @@ Summarization Instructions:
loadInquireStatus,
processAllRecordings,
processBatchRecordings,
+ reembedAllRecordings,
// Default Prompts methods
loadDefaultPrompt,