AI Filter & Rank Any Dataset

Find the records that actually match what you mean. Semantic filtering and ranking with source-grounded AI evidence.

Filter and rank any Apify dataset by meaning using natural language. No complex regular expressions, no fragile keyword search, and no external AI API keys required.

┌─────────────────┐     ┌──────────────────────┐     ┌────────────────────┐
│  Apify Dataset  │ ──► │ Natural Language     │ ──► │  Filtered & Ranked │
│  (Read-Only)    │     │ What you actually want│    │  with Source Proof │
└─────────────────┘     └──────────────────────┘     └────────────────────┘

---

Why Use This Actor?

Keyword search fails on real data:

• Searching for "independent dentist" still matches large corporate chains and dental drill suppliers.
• Searching for "backend engineer" matches generic full-stack roles where someone mentioned a backend framework once.
• Searching for "EV battery degradation" matches thousands of social posts that merely mention electric cars.

This Actor acts as an intelligent decision engine:

1. Understands Meaning: Compiles your instructions into an objective, calibrated scoring rubric with hard exclusions, mandatory requirements, and preference signals.
2. Grounds Every Judgment: Mechanically validates every AI citation against verbatim quotes in your source records. If the text does not say it, it is rejected.
3. Separates Facts from Inferences: Clearly distinguishes between observed source facts and analytical AI inferences.
4. Globally Calibrated & Reranked: High Quality mode evaluates records consistently across batches and performs required comparative listwise reranking over top candidates. If that required stage fails, the run fails without success output.
5. AI Included: Powered by self-hosted high-speed Chevet AI inference with automated reliability fallback to DeepSeek V4 Flash. You never need an external AI key.

---

Quick Start in 3 Steps

1. Select Dataset: Pick any existing Apify dataset (e.g. from Google Maps, LinkedIn, Twitter/X, Amazon, Web Scraper, or custom data).
2. Describe What You Want: Write instructions in plain English specifying your criteria, exclusions, and what to prioritize.
3. Run: Get a new clean dataset sorted with best matches first, complete with semantic scores (0–100) and verified source evidence.

---

Real-World Use Cases & Examples

1. B2B Lead Generation & Local Services

Dataset: Google Maps Scraper or Business Directory
Criteria:

"Find independent dental practices that appear to be real clinics rather than dental suppliers or national chains. Prioritize businesses where the available information suggests their website could benefit from modernization or lacks online booking."

2. Recruitment & Labour Market Intelligence

Dataset: Job Board or LinkedIn Scraper
Criteria:

"Find backend engineering jobs substantially focused on distributed systems, Go, Rust, or cloud infrastructure. Exclude generic full-stack positions, agency dev shops, or roles primarily focused on UI frameworks."

3. Social Media & Content Research

Dataset: TikTok, Instagram, or Twitter Scraper
Criteria:

"Find social media posts that genuinely discuss electric vehicle battery degradation, warranty claims, or long-term capacity loss. Exclude generic EV enthusiasm, unboxing videos, or posts merely mentioning Tesla or cars."

4. RAG & AI Knowledge Ingestion

Dataset: Scraped Web Pages or Documentation
Criteria:

"Keep only pages containing concrete architectural implementation details for OAuth token refresh flows, token rotation, and refresh token revocation. Exclude generic high-level marketing pages or basic introductory overviews."

5. Ecommerce & Product Fit Analysis

Dataset: Amazon or E-commerce Product Scraper
Criteria:

"Find portable camping power stations that explicitly support pass-through charging and LiFePO4 battery chemistry. Prioritize units weighing under 15 lbs with USB-C PD 100W output."

---

Output Contract & Transparency

Every record in the output dataset includes structured provenance and evidence:

{
  "rank": 1,
  "score": 94,
  "decision": "match",
  "modelConfidence": 0.91,
  "reason": "Independent family-run dental practice established in 1994, operating as a standalone local surgery.",
  "evidence": [
    {
      "path": "$.description",
      "kind": "quote",
      "quote": "Independent family-run dental practice established in 1994",
      "verified": true
    }
  ],
  "inferences": [
    {
      "statement": "The business appears to operate independently rather than as a corporate healthcare chain.",
      "support": ["e1"]
    }
  ],
  "analysis": {
    "provider": "chevet",
    "reviewed": true,
    "rankingMethod": "global_rerank",
    "contentCoverage": "full"
  },
  "source": {
    "datasetId": "w3K9m...2L",
    "itemIndex": 42,
    "recordId": "w3K9m...2L:42"
  },
  "item": {
    "name": "Oakfield Family Dental Care",
    "category": "Dental Clinic",
    "description": "Independent family-run dental practice established in 1994 by Dr. Sarah Jenkins...",
    "website": "http://oakfield-dentistry.co.uk"
  }
}

---

Security, Permissions & Privacy

• Limited Permissions: This Actor strictly requests READ permission on your source dataset. It never modifies, deletes, or mutates your original data.
• Prompt-Injection Resistance: Dataset records are treated strictly as inert data. Hostile instructions inside records (such as "IGNORE ALL PREVIOUS INSTRUCTIONS") are inert and cannot hijack the scoring rubric.
• Privacy Safe: Zero user records, personal emails, or scraped content are ever logged. Provider error bodies are not logged.
• Provider processing: The selected source fields are sent to ChevetAI Qwen for analysis and, only when necessary, to the disclosed OpenRouter DeepSeek fallback. Do not submit data you are not authorized to process.
• Original-record option: includeOriginalRecord copies the full source item into each output row. Disable it when the AI evidence and score are sufficient.

---

Pay-Per-Event (PPE) Monetization

This Actor is monetized fairly via Pay-Per-Event:

• Event: record-analyzed
• Fair Billing: You are charged strictly per successfully evaluated and validated source record.
• No Waste: Unread records, provider retries, internal repairs, or budget-halted records are never charged.
• Spending Limit Protection: If you set an Apify run spending limit, the Actor checks its budget and terminates cleanly when your limit is reached.
• AI work limits: maxAiCalls and maxAiTokens bound total provider work across evaluation, repair, review, adjudication, and reranking. A run with valid but incomplete coverage is explicitly labelled partial.
• Coverage transparency: SUMMARY.json reports requested, evaluated, failed, skipped, valid, full-content, and partial-content counts. A partial-content record is never presented as if its entire source record was evaluated.