Feature request: Instagram Posts (feed/carousel) as a source + WhisperX local transcription fallback for Reels + deep comment sentiment analysis

Feature Request: Instagram Posts as a Source + WhisperX Local Transcription + Deep Comment Sentiment Analysis

TL;DR

Expand the Instagram source beyond Reels to include Instagram Posts (feed posts, carousels, photo posts) via keyword search, add WhisperX as a local transcription fallback when cloud-based transcript APIs fail or are unavailable, and add deep comment sentiment analysis to surface what people actually think (not just what's most liked).

Problem

Currently, the Instagram source (scripts/lib/instagram.py) only searches Reels via two ScrapeCreators endpoints:

• /v2/instagram/reels/search (keyword search)
• /v1/instagram/user/reels (profile-based, --ig-creators)

This misses a massive chunk of Instagram's conversation:

1. Feed posts and carousels — where brands, creators, and communities share detailed opinions, product reviews, tutorials, and announcements. These often have more thoughtful comments than Reels.
2. Transcription reliability — the current pipeline relies on ScrapeCreators' /v2/instagram/media/transcript endpoint (#983 reports it frequently fails with Cloudflare 403s). There's no local fallback, so when the cloud transcript API is down, all spoken-word content from Reels is lost.
3. Comment sentiment — Instagram comments are fetched (PR #751 wired /v2/instagram/post/comments) but only used for ranking/engagement scoring. Comments are a goldmine of unfiltered opinion — people share detailed product experiences, political takes, emotional reactions — and this signal isn't being analyzed for sentiment or summarized into themes.

Proposed Changes

1. Instagram Posts as a Searchable Source

Add support for searching Instagram Posts (not just Reels) via ScrapeCreators or a complementary endpoint.

What's needed:

• A new search_instagram_posts() function (or extending search_instagram()) that hits a posts-specific ScrapeCreators endpoint (e.g., /v2/instagram/posts/search if available, or /v1/instagram/user/posts for profile-based fetching)
• _parse_items() already handles generic engagement fields (like_count, comment_count, caption) — may need minor adjustments for post-specific fields (carousel slide count, image URLs, multiple caption segments)
• A --ig-posts flag or automatic inclusion when INCLUDE_SOURCES=instagram covers both Reels and Posts
• Date filtering and deduplication (already handled by the merge logic in search_and_enrich())

Why this matters: Instagram Posts are where the detailed conversation happens. A product review carousel from a beauty creator can have 500+ comments with specific product feedback that never appears in a 15-second Reel. Brand announcements, long-form photo captions, and multi-slide tutorials live here — not in Reels.

Open question: Does ScrapeCreators expose a posts search endpoint? If not, the --ig-creators profile fetch could be extended to pull user posts alongside reels (similar to _user_reels() but for /v1/instagram/user/posts).

2. WhisperX as a Local Transcription Fallback

WhisperX is an open-source, locally-runnable speech recognition tool with:

• 70x realtime batched inference (large-v2 model)
• Word-level timestamps via wav2vec2 alignment
• Speaker diarization (multi-speaker identification via pyannote-audio)
• VAD preprocessing to reduce hallucinations
• Works on CPU (with --compute_type int8) or GPU

Integration approach:

Current flow:
  fetch_captions() → ScrapeCreators /v2/instagram/media/transcript → transcript text
  
Proposed flow:
  fetch_captions() → ScrapeCreators /v2/instagram/media/transcript → transcript text
                    ↓ (on failure/timeout/missing)
                    → Download audio from Reel URL (yt-dlp)
                    → Run WhisperX locally (whisperx audio.wav --model large-v2)
                    → Parse transcript segments + speaker labels
                    → Return transcript text + optional speaker-attributed segments

What's needed:

• A TranscriptionBackend abstraction in instagram.py (or a new transcription.py module):
  • CloudTranscriptionBackend — current ScrapeCreators /v2/instagram/media/transcript path
  • WhisperXBackend — local fallback using the whisperx Python package
• Config flag: TRANSCRIPTION_BACKEND=cloud|local|auto (default: auto = try cloud, fall back to local)
• Audio download via yt-dlp (already a dependency for YouTube)
• WhisperX model caching (first run downloads the model, subsequent runs reuse it)
• GPU detection with CPU fallback (--compute_type int8 --device cpu)
• Timeout handling (WhisperX on CPU can be slow for long videos — cap at 3 minutes of audio or configurable limit)

Why this matters:

• Issue #983 shows cloud transcription is unreliable (Cloudflare 403s, no User-Agent header)
• Issue #829 showed a single transcript 400 error discarding the entire result batch
• Local transcription is free after initial setup and works regardless of API availability
• WhisperX's speaker diarization is a unique value-add: knowing who said what in a multi-person Reel or interview clip would dramatically improve synthesis quality

Related: This would benefit YouTube and TikTok transcription too, not just Instagram. The TranscriptionBackend abstraction should be source-agnostic.

3. Deep Comment Sentiment Analysis

Currently, Instagram comments are fetched via /v2/instagram/post/comments and used for:

• Engagement scoring (comment_like_count → ranking signal)
• Display in "Top Community Comments" section (PR #751, rank-based diversity)

What's missing: actual analysis of what commenters are saying.

Proposed addition:

For each enriched post (top N by engagement):
  1. Cluster comments by theme/topic (e.g., "price complaints", "love the design", "shipping issues")
  2. Compute per-cluster sentiment (positive/negative/neutral) with confidence
  3. Surface representative quotes per cluster
  4. Generate a "Comment Sentiment Summary" per post:
     - Overall sentiment distribution (72% positive, 18% neutral, 10% negative)
     - Top 3 themes with representative comments
     - Controversial/nuanced takes (mixed-signal comments)
  5. Feed sentiment data into the synthesis: "Commenters on @creator's review post were 
     overwhelmingly positive about the build quality (82% of 340 comments) but 
     frustrated with the $299 price point"

Implementation approach:

• Use the existing LLM pass (the synthesis step) to cluster and analyze comments — no new model needed
• Add a sentiment_summary field to each enriched item
• Include comment sentiment in the brief's output format (new section or inline in item cards)
• For high-comment posts (>100 comments), sample representative comments across the sentiment distribution rather than just top-liked

Why this matters: Instagram comments are uniquely valuable because:

• People share detailed, unfiltered opinions (product reviews, emotional reactions, political hot takes)
• Comment volume is high (posts regularly get 200-2000+ comments)
• The audience is different from Reddit/X — more diverse demographics, more visual/brand-oriented
• Comment sentiment patterns (e.g., a product hyped in Reels but panned in post comments) can reveal gaps between marketing and reality

Scope & Phasing

Configuration Sketch

# .env or config
TRANSCRIPTION_BACKEND=auto          # cloud | local | auto
WHISPERX_MODEL=large-v2            # base, small, medium, large-v2
WHISPERX_DEVICE=auto               # auto (GPU if available) | cpu | cuda
WHISPERX_MAX_AUDIO_SECONDS=180     # Skip transcription for videos > 3 min

INCLUDE_SOURCES=tiktok,instagram,instagram_posts,youtube_comments,tiktok_comments,instagram_comments

References

• Current Instagram source: scripts/lib/instagram.py (636 lines)
• Instagram comments PR: #751 (enrich_with_comments → /v2/instagram/post/comments)
• Cloud transcription failures: #983 (Groq/Cloudflare 403), #829 (single 400 discards batch)
• Instagram Reels-only search issue: #957 (profile lane parsing), #986 (404 on keyword search)
• WhisperX: m-bain/whisperx — 70x realtime, word-level timestamps, speaker diarization