Skip to main content

Models Overview

Alactic AGI uses specialized AI models optimized for document intelligence and web content extraction. This guide explains the models, their capabilities, and how they were trained.

Available Models​

Alactic GPT-4o​

Full-featured model for complex document analysis

Base Model: OpenAI GPT-4o (June 2024 release)
Training Dataset: 1 billion+ tokens of document-specific content
Context Window: 128,000 tokens
Availability: Pro, Pro+, Enterprise plans

Optimizations:

  • Fine-tuned on structured document extraction
  • Enhanced table and chart understanding
  • Improved multi-page context retention
  • Better handling of technical terminology
  • Specialized prompting for information extraction

Best For:

  • Complex technical documents
  • Multi-page contracts and agreements
  • Financial reports with tables
  • Academic papers with citations
  • Legal documents requiring precision
  • Medical records and clinical notes

Performance:

  • Accuracy: 95%+ on standard document benchmarks
  • Speed: 50-100 tokens/second generation
  • Cost: $2.50 per 1M input tokens, $10.00 per 1M output tokens

Alactic GPT-4o mini​

Cost-effective model for routine document processing

Base Model: OpenAI GPT-4o mini
Training Dataset: 1 billion+ tokens of document-specific content
Context Window: 128,000 tokens
Availability: All plans (Free, Pro, Pro+, Enterprise)

Optimizations:

  • Fine-tuned on common document patterns
  • Efficient extraction of key-value pairs
  • Fast summarization capabilities
  • Optimized for web page content
  • Lower latency than full GPT-4o

Best For:

  • News articles and blog posts
  • Product descriptions
  • FAQ extraction
  • Simple forms and invoices
  • Email content analysis
  • Basic PDF text extraction

Performance:

  • Accuracy: 90%+ on standard document benchmarks
  • Speed: 100-150 tokens/second generation
  • Cost: $0.150 per 1M input tokens, $0.600 per 1M output tokens

Alactic Embedding Model​

Vector embedding for semantic search

Base Model: text-embedding-3-large
Dimensions: 3,072
Availability: All plans

Use Cases:

  • Semantic document search
  • Similarity detection
  • Content clustering
  • Duplicate identification
  • Related document suggestions

Performance:

  • Speed: Sub-second embedding generation
  • Cost: $0.130 per 1M tokens

Model Training Methodology​

Training Dataset​

Alactic models are fine-tuned on a proprietary dataset exceeding 1 billion tokens, carefully curated for document intelligence tasks.

Dataset Composition:

  • Public Documents: 300M tokens

    • Wikipedia articles
    • ArXiv research papers
    • GitHub documentation
    • Government public records
    • Open-access journals

  • Licensed Content: 400M tokens

    • Financial reports (10-K, 10-Q filings)
    • Legal contracts and agreements
    • Medical literature and case studies
    • Technical manuals and specifications
    • Academic textbooks

  • Synthetic Data: 300M tokens

    • Generated structured documents
    • Table and chart variations
    • Multi-format conversions
    • Edge case scenarios
    • Adversarial examples

Data Quality:

  • Manual verification of 10% sample
  • Automated quality checks
  • De-duplication at document level
  • Balanced representation across domains
  • Regular updates with new content

Fine-Tuning Process​

Stage 1: Domain Adaptation (GPT-4o base)

  • Continued pre-training on document corpus
  • Learning document-specific patterns
  • Vocabulary expansion for technical terms
  • Duration: 100,000 training steps

Stage 2: Instruction Tuning

  • Task-specific instruction datasets
  • Extract, summarize, analyze, compare tasks
  • Multi-turn conversation examples
  • Duration: 50,000 training steps

Stage 3: Reinforcement Learning from Human Feedback (RLHF)

  • Human evaluation of model outputs
  • Preference ranking of responses
  • Reward model training
  • Policy optimization
  • Duration: 20,000 training steps

Total Training: 170,000 steps over 4 weeks on Azure ML infrastructure

Evaluation Benchmarks​

Models are evaluated against industry-standard benchmarks:

Document Understanding:

  • DocVQA (Document Visual Question Answering): 92.3% accuracy
  • InfoVQA (Infographic Question Answering): 89.7% accuracy
  • TabFact (Table Fact Verification): 94.1% accuracy

Information Extraction:

  • FUNSD (Form Understanding): 91.5% F1 score
  • CORD (Consolidated Receipt Dataset): 96.2% F1 score
  • SROIE (Scanned Receipt OCR): 97.8% F1 score

Summarization:

  • CNN/DailyMail: 45.2 ROUGE-L score
  • XSum: 48.7 ROUGE-L score
  • PubMed: 43.9 ROUGE-L score

Web Scraping:

  • Common Crawl extraction: 93.4% precision
  • News article parsing: 96.1% precision
  • E-commerce product data: 94.8% precision

Model Capabilities Comparison​

CapabilityGPT-4oGPT-4o mini
Text ExtractionExcellentExcellent
Table ExtractionExcellentGood
Chart UnderstandingExcellentFair
Multi-page ContextExcellentGood
Technical TermsExcellentGood
Summarization QualityExcellentGood
Reasoning DepthDeepModerate
SpeedFastFaster
CostHigherLower

When to Use Which Model​

Use GPT-4o for:​

  1. Complex Documents

    • Multi-page contracts
    • Financial statements
    • Research papers
    • Technical manuals

  2. High Accuracy Requirements

    • Legal document review
    • Medical record analysis
    • Compliance checking
    • Audit trails

  3. Deep Analysis

    • Comparative analysis
    • Trend identification
    • Risk assessment
    • Strategic insights

  4. Multi-step Reasoning

    • Chain-of-thought processing
    • Multi-document synthesis
    • Causal relationship detection

Use GPT-4o mini for:​

  1. Simple Documents

    • News articles
    • Blog posts
    • Product descriptions
    • Email content

  2. High-Volume Processing

    • Batch document processing
    • Real-time scraping
    • Routine data extraction
    • Classification tasks

  3. Cost Optimization

    • Large document sets
    • Frequent processing
    • Development and testing
    • Non-critical applications

  4. Speed-Critical Applications

    • Real-time analysis
    • Interactive applications
    • Low-latency requirements

Model Selection in Application​

Automatic Selection​

Alactic AGI automatically selects the appropriate model based on:

  • Your plan tier
  • Document complexity
  • Processing volume
  • Cost optimization settings

Free Plan: Always uses GPT-4o mini
Pro/Pro+/Enterprise: User can choose or use auto-selection

Manual Override​

Users on Pro+ and Enterprise plans can force specific model usage:

{
  "document": "complex-contract.pdf",
  "model": "gpt-4o",
  "force": true
}

Smart Routing (Enterprise Only)​

Enterprise deployments can enable smart routing:

  • Simple documents → GPT-4o mini automatically
  • Complex documents → GPT-4o automatically
  • Based on machine learning classifier
  • Reduces costs by 40-60% typically

Token Usage and Costs​

Input Token Calculation​

Documents:

  • Text-only PDF: ~250 tokens per page
  • PDF with tables: ~400 tokens per page
  • Scanned PDF (OCR): ~300 tokens per page

Web Pages:

  • News article: 500-1,500 tokens
  • Blog post: 800-2,000 tokens
  • Product page: 300-800 tokens
  • Documentation page: 1,000-3,000 tokens

Output Token Calculation​

Typical Outputs:

  • Summary: 150-300 tokens
  • Key points extraction: 100-200 tokens
  • Full analysis: 500-1,000 tokens
  • Structured data: 200-400 tokens

Cost Examples​

Processing 100 PDFs (10 pages each) with GPT-4o:

Input: 100 docs × 10 pages × 400 tokens = 400,000 tokens = 0.4M tokens
Output: 100 docs × 300 tokens = 30,000 tokens = 0.03M tokens

Cost: (0.4M × $2.50) + (0.03M × $10.00) = $1.00 + $0.30 = $1.30

Processing 100 PDFs (10 pages each) with GPT-4o mini:

Input: 400,000 tokens = 0.4M tokens
Output: 30,000 tokens = 0.03M tokens

Cost: (0.4M × $0.150) + (0.03M × $0.600) = $0.06 + $0.018 = $0.078

Savings: 94% cost reduction using GPT-4o mini for suitable documents

Model Performance Optimization​

Prompt Engineering​

Alactic AGI uses optimized prompts for each task:

  • Extraction prompts include examples
  • Summarization prompts specify format
  • Analysis prompts provide context
  • Structured output uses JSON mode

Context Management​

  • Automatic chunking for large documents
  • Overlapping context windows
  • Smart section selection
  • Page reference preservation

Caching Strategy​

  • Common prompts are cached
  • Reduces redundant token usage
  • Up to 50% token savings
  • Automatic cache invalidation

Batch Processing​

  • Documents processed in optimal batches
  • Parallel API calls when possible
  • Rate limit management
  • Error handling and retries

Model Updates and Versioning​

Version Numbering​

Format: alactic-{model}-{version}

Examples:

  • alactic-gpt4o-v1.0: Initial release
  • alactic-gpt4o-v1.1: Bug fixes and improvements
  • alactic-gpt4o-v2.0: Major update with new training

Update Frequency​

  • Minor updates: Monthly (bug fixes, optimizations)
  • Major updates: Quarterly (new training data, capabilities)
  • Hotfixes: As needed (critical issues)

Backwards Compatibility​

  • API contracts remain stable
  • Output format consistency maintained
  • Gradual deprecation of old versions (6-month notice)

Staying Updated​

  • Release notes published at /docs/release-notes
  • Email notifications for Enterprise customers
  • Optional auto-update (default: enabled)
  • Pinning specific versions available

Future Model Roadmap​

Q1 2026​

  • Multimodal Support: Direct image and chart analysis
  • Longer Context: 200K token context window
  • Faster Processing: 2x speed improvement

Q2 2026​

  • Custom Model Training: Upload your own training data
  • Domain-Specific Models: Legal, medical, financial variants
  • Real-time Streaming: Token-by-token output

Q3 2026​

  • Multilingual Support: 50+ languages
  • Audio Transcription: Built-in speech-to-text
  • Video Analysis: Extract text and insights from videos

Q4 2026​

  • AGI Features: Multi-agent workflows
  • Reasoning Models: Enhanced logical reasoning
  • Tool Use: Models can call external APIs