Files
ALwrity/backend/api/podcast/cost_estimator.py
ajaysi 3f984e8d0c feat(podcast): add pre-estimate endpoint, enhance cost estimator with multi-model support, cleanup alpha pricing seeding
- Add POST /podcast/pre-estimate endpoint for cost estimation before analysis
- Enhance cost_estimator.py with multi-model support (gemini, audio, voice clone, image, video)
- Add detailed cost breakdown (llm, audio, media costs + per-phase breakdown)
- Remove redundant pricing seeding from init_alpha_subscription_tiers.py
- Add SSOT pricing via PricingService.initialize_default_pricing()
- Update TopicUrlInput tooltip to show estimate details
- Add debug logging for pricing seeding and pre-estimate
- Clean up verbose podcast mode debug logs in app.py
2026-05-06 15:29:12 +05:30

216 lines
7.9 KiB
Python

"""
Podcast cost estimation helpers.
Builds user-facing podcast estimates from the subscription pricing catalog
instead of hard-coded frontend heuristics.
Supports multiple models for each component:
- Audio TTS: minimax/speech-02-hd (default), qwen3-tts, cosyvoice-tts
- Voice Clone: qwen3, cosyvoice, minimax
- Image: qwen-image (default), ideogram-v3-turbo
- Video: wan-2.5 (default), kling-v2.5, infinitetalk
- LLM: gemini-2.5-flash (default)
"""
from __future__ import annotations
from typing import Any, Dict, Optional
from sqlalchemy.orm import Session
from models.subscription_models import APIProvider
from services.subscription.pricing_service import PricingService
def _round_money(value: float) -> float:
return round(float(value), 4)
def _load_pricing(
pricing_service: PricingService,
provider: APIProvider,
preferred_model: str,
) -> Optional[Dict[str, Any]]:
"""Load pricing for a provider and model, with fallback to default."""
pricing = pricing_service.get_pricing_for_provider_model(provider, preferred_model)
if pricing:
return pricing
# Fallback to provider default model row (if configured).
return pricing_service.get_pricing_for_provider_model(provider, "default")
# Default models used in podcast generation
DEFAULT_MODELS = {
"gemini": "gemini-2.5-flash",
"exa": "exa-search",
"audio_tts": "minimax/speech-02-hd",
"voice_clone": "wavespeed-ai/qwen3-tts/voice-clone",
"image": "qwen-image",
"video": "wan-2.5",
}
def estimate_podcast_cost(
*,
db: Session,
duration_minutes: int,
speakers: int,
query_count: int,
include_avatar_phase: bool = True,
# Optional model overrides
gemini_model: str = "gemini-2.5-flash",
audio_tts_model: str = "minimax/speech-02-hd",
voice_clone_engine: str = "qwen3",
image_model: str = "qwen-image",
video_model: str = "wan-2.5",
) -> Optional[Dict[str, Any]]:
"""
Compute a backend estimate for podcast creation.
Supports customizable models for each component.
Uses pricing_catalog for accurate cost calculation.
"""
pricing_service = PricingService(db)
# Load pricing for each component and model
gemini_pricing = _load_pricing(pricing_service, APIProvider.GEMINI, gemini_model)
exa_pricing = _load_pricing(pricing_service, APIProvider.EXA, "exa-search")
# Audio TTS pricing (minimax/speech-02-hd)
audio_pricing = _load_pricing(pricing_service, APIProvider.AUDIO, audio_tts_model)
# Voice clone pricing (different engines)
voice_clone_model = f"wavespeed-ai/{voice_clone_engine}-tts/voice-clone"
voice_clone_pricing = _load_pricing(pricing_service, APIProvider.AUDIO, voice_clone_model)
if not voice_clone_pricing:
# Try alternate model names
voice_clone_pricing = _load_pricing(pricing_service, APIProvider.AUDIO, f"{voice_clone_engine}/voice-clone")
# Image pricing (qwen-image or ideogram)
image_pricing = _load_pricing(pricing_service, APIProvider.STABILITY, image_model)
# Video pricing (wan-2.5, kling, or infinitetalk)
video_pricing = _load_pricing(pricing_service, APIProvider.VIDEO, video_model)
# Return None if critical pricing unavailable (fail fast)
if not gemini_pricing:
return None
# Configuration
minutes = max(1, int(duration_minutes or 1))
speaker_count = max(1, int(speakers or 1))
research_queries = max(1, int(query_count or 1))
# Token usage assumptions per phase
analysis_input_tokens = 1800
analysis_output_tokens = 1000
research_synthesis_input_tokens = 2200
research_synthesis_output_tokens = 900
script_input_tokens = max(1800, minutes * 300)
script_output_tokens = max(2200, minutes * 700)
# TTS: ~900 chars per minute per speaker
estimated_tts_tokens = max(900, minutes * 900 * speaker_count)
# Voice clone: 1 clone operation per speaker
voice_clone_count = speaker_count
# ===== COST CALCULATIONS =====
# 1. Analysis phase (LLM)
analysis_cost = (
analysis_input_tokens * float(gemini_pricing.get("cost_per_input_token") or 0.0)
+ analysis_output_tokens * float(gemini_pricing.get("cost_per_output_token") or 0.0)
)
# 2. Research phase
# 2a. LLM for research synthesis
research_llm_cost = (
research_synthesis_input_tokens * float(gemini_pricing.get("cost_per_input_token") or 0.0)
+ research_synthesis_output_tokens * float(gemini_pricing.get("cost_per_output_token") or 0.0)
)
# 2b. Search API (Exa)
research_search_cost = 0.0
if exa_pricing:
research_search_cost = research_queries * float(exa_pricing.get("cost_per_request") or 0.0)
research_cost = research_search_cost + research_llm_cost
# 3. Script generation (LLM)
script_cost = (
script_input_tokens * float(gemini_pricing.get("cost_per_input_token") or 0.0)
+ script_output_tokens * float(gemini_pricing.get("cost_per_output_token") or 0.0)
)
# 4. Audio TTS
tts_cost = 0.0
if audio_pricing:
tts_cost = estimated_tts_tokens * float(audio_pricing.get("cost_per_input_token") or 0.0)
# 5. Voice cloning (if needed)
voice_clone_cost = 0.0
if voice_clone_pricing:
voice_clone_cost = voice_clone_count * (
float(voice_clone_pricing.get("cost_per_request") or 0.0)
+ estimated_tts_tokens * float(voice_clone_pricing.get("cost_per_input_token") or 0.0)
)
# 6. Avatar image generation
avatar_cost = 0.0
if include_avatar_phase and image_pricing:
image_unit = float(image_pricing.get("cost_per_image") or image_pricing.get("cost_per_request") or 0.0)
avatar_cost = speaker_count * image_unit
# 7. Video rendering
video_cost = 0.0
if video_pricing:
# Assume 1 video render per minute (upper bound)
video_cost = minutes * float(video_pricing.get("cost_per_request") or 0.0)
# ===== TOTALS =====
llm_total = analysis_cost + research_llm_cost + script_cost
audio_total = tts_cost + voice_clone_cost
media_total = avatar_cost + video_cost
total = llm_total + research_search_cost + audio_total + media_total
return {
# Cost breakdown
"analysisCost": _round_money(analysis_cost),
"researchCost": _round_money(research_cost),
"researchSearchCost": _round_money(research_search_cost),
"researchLlmCost": _round_money(research_llm_cost),
"scriptCost": _round_money(script_cost),
"ttsCost": _round_money(tts_cost),
"voiceCloneCost": _round_money(voice_clone_cost),
"avatarCost": _round_money(avatar_cost),
"videoCost": _round_money(video_cost),
"total": _round_money(total),
# Totals by category
"llmCost": _round_money(llm_total),
"audioCost": _round_money(audio_total),
"mediaCost": _round_money(media_total),
# Currency
"currency": "USD",
"source": "pricing_catalog",
# Models used for this estimate
"models": {
"llm": gemini_model,
"research": "exa-search",
"audio_tts": audio_tts_model,
"voice_clone": voice_clone_model,
"image": image_model,
"video": video_model,
},
# Assumptions used
"assumptions": {
"analysis_input_tokens": analysis_input_tokens,
"analysis_output_tokens": analysis_output_tokens,
"research_synthesis_input_tokens": research_synthesis_input_tokens,
"research_synthesis_output_tokens": research_synthesis_output_tokens,
"script_input_tokens": script_input_tokens,
"script_output_tokens": script_output_tokens,
"estimated_tts_tokens": estimated_tts_tokens,
"research_queries": research_queries,
"voice_clone_count": voice_clone_count,
"video_requests": minutes,
"avatar_requests": speaker_count if include_avatar_phase else 0,
},
}