Spaces:
Sleeping
Sleeping
Commit History
-
fix: normalize search_volume (int→str), relevance (float→int), competition (lower→Title), post_count (int→str) for LongCat API responses
3ef98d0 -
fix: add catch-all viral_hashtags mapping at end of cross_map_fields to catch hashtags set by platform-specific mappers
2680851 -
fix: add generic cross-map fields (core_keywords, viral_hashtags, strategy_tips, content_titles) to all platform response models so they survive FastAPI's response_model filter
91c8715 -
fix: add Pinterest cross-maps for core_keywords, content_titles, hashtags, strategy_tips
504cc76 -
fix: handle model returning dict instead of list for tweet_threads/post_drafts/post_ideas/video_concepts
468fc59 -
feat: Add LinkedIn API posting + enhanced analyzer + schema
639b959 -
feat: Add LongCat API as primary model with Qwen fallback + cross-map fix
249b12a -
fix: Add Instagram cross-maps for hashtag_sets/captions/content_ideas
43bad80 -
fix: Run cross-map after defaults so LLM-sparse responses still get mapped
0054399 -
fix: Add core_keywords and related_phrases cross-mapping from tags/video_titles
2eb9ab6 -
fix: Cross-map platform keys to generic keys for frontend template
ae74dac -
fix: Improve prompts to return JSON object, add error for list response
5b10e71 -
fix: Add error message when LLM returns list instead of dict
a4fce70 -
fix: Escape curly braces in f-string causing Invalid format specifier
c19bbb8 -
fix: Guard against LLM returning JSON array instead of object
2cf0f7b -
fix: Add missing validate_and_fill_data_defaults function
758a047 -
fix: Catch normalize/validate errors, hide raw JSON parse errors againe
e22a390 -
fix: Catch normalize/validate errors, hide raw JSON parse errors
4a774c4 -
feat: Multi-platform batch generation + simple UI
- New /analyze/batch endpoint generates content for multiple platforms at once - New BatchRequest schema and batch_analyzer service - Frontend: checkbox multi-select, per-section copy buttons, platform tabs91bd745 -
feat: Simple HTML frontend with section copy buttons
Replace JSON root with static HTML UI. Each section has per-item and Copy All buttons. Supports all 7 SEO platforms + grammar. Removed View Raw JSON option.a1f2a19 -
fix: Handle missing correction fields + ResponseValidationError
c4fea6f -
feat: Add Grammar Correction API v2.1.0
- New POST /analyze/grammar endpoint (GrammarRequest/GrammarResponse schemas) - New grammar_analyzer.py with temperature 0.2, max_new_tokens 1200 - 11-field response: corrections, issues, grammar_score, readability, tone, etc. - Updated utils.py normalize_response() to handle grammar field types - Removed emojis from logs (model_loader.py, utils.py) for Windows cp1252 compat - Bumped version 2.0.0 -> 2.1.0, updated API title/descriptionff08af5
-
fix: Add normalize_response to handle model output format mismatches
- Model returns strings where objects expected (e.g., video_titles: ['str'] → [{'title':'str','expected_ctr':'Medium'}]) - Model returns objects where strings expected (e.g., thumbnail_ideas: [{url, caption}] → ['caption text']) - Normalizes all field types: keyword_list, hashtag_list, caption_list, pin_list, object_list, string_list - Handles nested structures: hashtag_sets, content_ideas, post_drafts, tweet_threadse1cb6ad -
perf: Reduce max_new_tokens 1500→800, cut output by 50% for faster CPU inference
- max_new_tokens: 1500 → 800 (cuts response time ~50%) - Output items reduced: 5→3 titles, 10→5 tags, 3→2 ideas - Topic truncation: 2000→1000 chars (less prompt processing) - Content truncation: 2000→1000 chars (less prompt processing) - Users on GPU can increase via .env configb0fb0e7 -
fix: Raise HTTPException for unexpected errors instead of returning invalid dict
- safe_analyze now raises HTTPException(500) on unexpected exceptions - This prevents Pydantic ResponseValidationError from missing fields - Expected errors (JSON parse fail) still return 200 with defaults + error0ca65c8 -
feat: Add 3-tier quantization with auto hardware detection
- Tier 1: Qwen2.5-7B with 4-bit NF4 quantization (needs GPU 5.5GB+) - Tier 2: Qwen2.5-1.5B with 8-bit quantization (GPU or CPU) - Tier 3: Qwen2.5-0.5B bfloat16 fallback (CPU, default) - Auto-detect GPU/CPU and select optimal quantization - Auto-fallback: 7B→1.5B if GPU VRAM insufficient - Configurable via .env: QUANTIZATION=auto/4bit/8bit/none - Graceful fallback if bitsandbytes not installed - Update requirements.txt with bitsandbytes>=0.43.0 - Update Dockerfile with GPU base image optionefb4fe4 -
fix: Simplify prompts for 0.5B model reliability + resilient error handling
- Reduce output requirements (5 titles, 10 tags, 3 thumbnails, etc.) - Add v2 JSON repair logic for truncated output - Always return defaults-filled data instead of 500 errors - Pass defaults to run_analysis for error resilience - Remove HTTPException 500 — API always returns 200 with error field6e467d1
-
feat: Add multi-platform social media analysis with 7 platform-specific endpoints
- Add 7 new platform analyzers: YouTube, Instagram, LinkedIn, Facebook, X/Twitter, TikTok, Pinterest - Create shared model loader with .env configurable MODEL_ID (0.5B/1.5B/7B) - Refactor services with shared JSON repair and validation utilities - Update schemas with 7 platform-specific Pydantic response models - Add temperature tuning per platform (precise 0.3 to creative 0.45) - Implement professional error handling with safe_analyze wrapper - Add .env.example for model configuration documentation - Keep backward-compatible original /analyze-seo route38098b4
-
Add GPU Setup Cheat Sheet for RTX 4050
30aa031 -
Add documentation for running on RTX 4050 6GB
2caec2a -
Update docs to reflect Qwen-0.5B Local CPU architecture
bb2afa9 -
Add Schema Validation Safety Net: Auto-fill missing JSON fields with defaults
c31af39 -
Fix JSON Crash: Lower Temp to 0.3 + Enhance Repair Logic + Add Error Fallback
fdbd75a -
Switch to Qwen-0.5B (Nano) for instant speed and 100% stability
f169fdd -
Optimize for Speed: 30 Keywords / 1500 Tokens to stop Local CPU Timeout
99776a0 -
Reduce to 50 keywords to prevent CPU Timeout (Internal Server Error)
4d779c6 -
Scale up to 150 keywords using Local Inference
9a256bc -
Migrate to Local Inference (Qwen-1.5B) to permanently bypass API limits: Re-added Torch/Transformers
50196f5 -
Switch to SmolLM2 (HF Native) to resolve provider support issues
e7e750b -
Switch to Qwen-1.5B (Lightweight) to fix all API limits
24ee997 -
Switch to Google Gemma-2-9B-IT for dependable serverless support
ddcd3d9 -
Fix NameError and Switch to Phi-3-Mini to bypass 402 Payment Limits
3509502 -
Fix truncation error: Reduce to 50 keywords + JSON repair logic
7252943 -
Increase keyword limit to 150 and max_tokens to 8000
394498e -
Extend output: 10 keywords/hashtags + add Content Titles
e8267ad -
Update docs with A-Z guide for Qwen Model
d68c4e0 -
Update prompts for stronger SEO data extraction
3d17470 -
Switch to Qwen/Qwen2.5-7B-Instruct for reliable availability
1b66ca1 -
Switch to Zephyr-7B-Beta for Chat API compatibility
ad41aff -
Fix API method: Use chat_completion for Mistral Instruct
d877e39