Social Comment Buyer Signal Filter

Turn already-collected Instagram, TikTok, YouTube and Threads comments into auditable buyer and customer-service signals. Import an existing Apify Dataset or supply JSON. Get purchase-location questions, price inquiries, shipping questions, product questions, complaints, refund/return requests and general comments.

What it does

• Normalizes common comment fields and preserves explicit platform values.
• Returns matched phrases, all matching categories, a primary label and a readable reason.
• Skips empty/invalid comments and deduplicates within a run.
• Optionally outputs only signals or omits raw text.
• Uses deterministic English/Korean phrase rules. No external AI, scraping, proxy or API key is required for inline input.

This is a triage tool, not a prediction of purchases. It does not contact commenters, verify demand, or scrape any social network. Sarcasm, short context-dependent comments and unsupported languages may need review. It is not a general multilingual sentiment model.

Quick start

Set items to an array of objects or strings, then Start. The example in example-input.json produces four rows: Instagram/BUY_QUESTION, TikTok/SHIPPING_QUESTION, YouTube/PRODUCT_COMPLAINT, Threads/GENERAL_COMMENT.

For a previously collected dataset, set sourceDataset to its ID. The run's Apify identity must have read access. Dataset input replaces inline input, preventing accidental mixing of demonstration rows. maxItems limits scanned rows, including duplicates and empty rows; maximum 100,000.

Input

Field	Default	Meaning
items	No runtime default	Array of comments or strings
sourceDataset	Empty	Existing Dataset ID; takes priority
maxItems	1000	Maximum source rows, 1–100000
platformHint	auto	Explicit manual override or auto
signalsOnly	false	Omit GENERAL_COMMENT rows
includeText	true	False omits raw text; matched phrases remain
trackedTerms	Empty	Comma-separated optional tags, maximum 25

Text keys: text, comment, text_content, commentText, textOriginal, textDisplay, body, message. YouTube API snippet/topLevelComment.snippet text fields are also supported. Each input row represents one comment; nested reply arrays are not expanded.

Platform precedence: manual override, explicit platform/sourcePlatform, recognized URL hostname, platform-specific field combinations, unknown. Generic comment/commentUrl fields do not prove a platform. Conflicting explicit/URL evidence is flagged. Platform unknown does not itself make an otherwise clear text classification ambiguous.

Output

Field	Meaning
id	Source ID, or stable hash when absent
platform / platformReason / platformConflict	Platform and its evidence
text / textTruncated	First 5000 characters or empty when omitted
sourceUrl	First available source link
label / labels	Primary and all detected categories
matchedTerms / reason	Phrase evidence and explanation
needsReview	Ambiguous unmatched text, negation, truncation or platform conflict
ruleStrength	PHRASE_MATCH or NONE; not a calibrated probability
trackedTerms	Matching optional product tags
method / modelVersion	RULE_BASED / 3.0.0

Categories: BUY_QUESTION, PRICE_QUESTION, SHIPPING_QUESTION, PRODUCT_QUESTION, PRODUCT_COMPLAINT, RETURN_REFUND_QUESTION, GENERAL_COMMENT. Primary priority: return/refund, complaint, buying, price, shipping, product. Multiple matches remain in labels; they do not automatically require review. Familiar general engagement such as “Great video!” does not automatically require review.

Deduplication key is platform + source URL + source ID. Without ID, a hash of platform, URL and full text is used. Distinct people posting identical text without IDs cannot be distinguished. Supply original IDs where possible.

SUMMARY in key-value storage reports scanned, invalid, empty, duplicate, processed, filtered and output counts, primary-category counts and spending-limit state. An input with no data raises an error. An input containing only empty, duplicate or filtered comments may legitimately produce zero results; inspect SUMMARY.

Privacy and reliability

Only normalized output fields are stored; author profiles are not copied. Input text and matched phrases may still contain personal data. Use only data you are authorized to process and manage source/input storage retention separately. No external text-analysis service receives comments.

The entry point is python -m social_filter; __main__.py explicitly runs the async main function. The SDK version is pinned. Every release should pass entry-point, classification, deduplication, cloud Dataset and spend-limit checks.

Pricing

Pay per event: $0.50 per 1,000 delivered comments ($0.0005 each), plus $0.001 per run for the Actor-start event at the allowed memory sizes. Platform usage is included. No subscription is required by this Actor.

The custom comment-analyzed event is charged once per delivered output row. Empty, invalid, duplicate and signal-filtered rows do not incur a result event. The start fee still applies when no results are produced. The automatic dataset-item event is disabled to prevent double billing.

Examples: four output rows cost $0.003 including startup; 1,000 output rows cost $0.501. A run limit of $0.002 allows startup plus two rows. The minimum selectable run limit is $0.0015 (startup plus one row). Set your own maximum charge before starting; the Actor stops output when the remaining budget cannot cover another row.

Cloud validation on 2026-11-03 verified two delivered rows at a $0.002 limit and four at a $0.01 limit with PPE enabled. These were developer-owned runs, not independent customer payment or payout validation.

Example input

{
  "items": [
    {"id": "sample-1", "platform": "instagram", "text": "Where can I buy this?"},
    {"id": "sample-2", "platform": "tiktok", "text": "Do you ship to Canada?"},
    {"id": "sample-3", "platform": "youtube", "text": "Mine broke after a week."},
    {"id": "sample-4", "platform": "threads", "text": "Great video!"}
  ],
  "maxItems": 1000,
  "signalsOnly": false
}

The four rows produce BUY_QUESTION, SHIPPING_QUESTION, PRODUCT_COMPLAINT and GENERAL_COMMENT respectively, with needsReview: false for these explicit examples. For MCP clients, use comment objects rather than raw strings because generated client schemas may require objects.