Backend AI Functions

​

Overview

• Extract structured data from conversations and documents
• Generate case insights and portfolio analytics
• Manage CaseMind™ memory and fact storage
• Auto-complete tasks and archive deadlines
• Track billable time and AI usage
• Orchestrate complex multi-step workflows

​

Case Intelligence

​

Case Insights (victoria-case-manager)

• Portfolio Health Scoring: Evaluates overall case load health
• Momentum Detection: Identifies stalled cases (no activity in 14+ days)
• Deadline Alerts: Highlights upcoming hearings and critical dates
• Phase Transition Recommendations: Suggests when cases are ready to move forward
• Documentation Gap Analysis: Detects missing financial documents or discovery materials

1. Runs automatically on dashboard load
2. 24-hour caching for performance
3. Analyzes all active cases in your firm
4. Generates 5-10 actionable insights with severity ratings (critical/high/medium/low)

{
  "type": "momentum",
  "severity": "high",
  "title": "Case stalled: Johnson v. Johnson",
  "description": "No activity in 18 days. Last action: Filed financial affidavit",
  "action_required": "Schedule status conference or follow up with opposing counsel"
}

​

CaseMind™ Memory System

​

Memory Extract (victoria-memory-extract)

• Financial: Income, assets, debts, property values
• Procedural: Court dates, filing deadlines, case status
• Strategic: Attorney recommendations, settlement positions
• Personal: Children’s ages, custody preferences, special needs
• Timeline: Key events, separation dates, incidents
• Discovery: Document requests, production deadlines
• Communication: Important client statements, opposing party positions

• Deduplication: Prevents redundant facts (70% similarity threshold)
• Entity Normalization: Standardizes dates, amounts, names
• Confidence Scoring: 0-100 score for fact reliability
• Automatic Conflict Detection: Flags contradictory information
• Retry Logic: Resilient to API failures

1. Triggers automatically after Victoria chat responses
2. Claude Haiku 4.5 analyzes conversation for extractable facts
3. Stores facts in victoria_case_memory table with metadata
4. Makes facts available to all Victoria agents for future context

​

Memory Load (victoria-memory-load)

• Filters facts by case_id and relevance
• Prioritizes recent facts (last 90 days)
• Includes critical facts regardless of age
• Excludes superseded facts (marked inactive by Lifecycle Manager)
• Formats facts for optimal LLM comprehension

CASEMIND MEMORY (Johnson v. Johnson):
• Husband's annual income: $125,000 (W-2, verified)
• Wife's annual income: $48,000 (self-employed, tax returns)
• Marital home value: $650,000 (appraisal dated 2024-10-15)
• Outstanding mortgage: $280,000
• 2 minor children: Emma (age 12), Lucas (age 9)
• Current custody: Wife has primary, husband weekends
• Separation date: 2024-03-15

​

Background Extraction & Orchestration

​

Background Orchestrator (victoria-background-orchestrator)

1. case_metadata: Extract case numbers, party names, judges, courts
2. casemind_facts: Extract legal facts and timeline events
3. deadlines: Extract court dates and filing deadlines
4. tasks: Extract action items and next steps
5. embeddings: Generate vector embeddings for semantic search

• Critical: Process immediately (e.g., deadline extraction)
• High: Process within 5 minutes (e.g., case metadata)
• Medium: Process within 15 minutes (e.g., CaseMind facts)
• Low: Process within 1 hour (e.g., embeddings)

• Per-Firm Budgets: Monthly extraction quotas and cost limits
• Deduplication Cache: 24-hour TTL prevents duplicate jobs
• Circuit Breakers: Auto-pause on API rate limits
• Retry Logic: Exponential backoff (3 max attempts)
• Cost Tracking: Real-time spend monitoring per firm

​

Extraction Worker (victoria-extraction-worker)

• Runs every 2-5 minutes via Supabase Cron
• Processes up to 50 jobs per run (configurable batch size)
• Calls Background Orchestrator to execute jobs
• Reports metrics: processed count, queue size, processing time

SELECT
  status,
  COUNT(*) as count,
  AVG(EXTRACT(EPOCH FROM (completed_at - created_at))) as avg_seconds
FROM background_extraction_jobs
WHERE created_at > NOW() - INTERVAL '24 hours'
GROUP BY status;

​

Specialized Extractors

​

Task Extractor (victoria-task-extractor)

• Looks for phrases like “you should,” “next step,” “make sure to”
• Identifies explicit action items in attorney-Victoria chats
• Extracts due dates and assigns to appropriate users
• Categorizes tasks: discovery, filing, client communication, investigation

Title: "File Motion to Modify Support"
Description: "Victoria recommended filing motion due to husband's income change ($125K → $95K)"
Due Date: 2025-02-15 (30 days from conversation)
Assigned To: Attorney handling case
Priority: High

​

Deadline Extractor (victoria-deadline-extractor)

• Court hearings and trials
• Discovery cutoffs (interrogatories, requests for production)
• Financial disclosure deadlines
• Response deadlines (motions, pleadings)
• Mediation and settlement conferences

• Recognizes jurisdiction-specific rules (e.g., Florida’s 45-day financial disclosure)
• Calculates response deadlines automatically
• Creates calendar events with reminders
• Flags conflicts with existing deadlines

​

Case Metadata Extract (victoria-case-metadata-extract)

• Case Number: Court-assigned case identifier
• Parties: Petitioner, respondent, children names/ages
• Court Information: Judge name, courtroom, jurisdiction
• Case Type: Dissolution, modification, contempt, paternity
• Attorney Information: Opposing counsel, their firm

​

Financial Affidavit Extract (victoria-financial-affidavit-extract)

• Income (W-2, self-employment, bonuses, etc.)
• Assets (bank accounts, investments, real property, vehicles)
• Liabilities (mortgages, credit cards, loans)
• Monthly expenses (housing, utilities, childcare, etc.)

​

Lifecycle Management

​

Lifecycle Manager (victoria-lifecycle-manager)

• Compares pending tasks to recent facts (last 7 days)
• 70% similarity threshold for auto-completion
• Marks task as completed with source fact reference

• Task: “File financial affidavit”
• Fact Detected: “Uploaded Florida Financial Affidavit (Form 12.902(c)) on 2025-01-15”
• Result: Task auto-completed ✅

• Asset value updates
• Income changes
• Address changes

​

Routing & Coordination

​

Orchestrator (victoria-orchestrator)

1. Analyzes user question with Claude Haiku 4.5 (fast, cheap)
2. Detects intent: legal strategy → Co-Counsel, financial → Financial Analyst, discovery → Discovery Manager
3. Routes to specialized agent with full context
4. Returns response to user

• Uses Haiku 4.5 for routing (~$0.002 per request)
• Prevents unnecessary expensive model calls
• ~95% routing accuracy

​

Question Router (victoria-question-router)

• General questions → Victoria Client Liaison (e.g., “What is community property?”)
• Case-specific legal advice → Attorney escalation (e.g., “Should I accept this settlement offer?”)
• Sensitive issues → Attorney escalation (e.g., domestic violence, complex custody)

​

Utility Functions

​

Context Injector (victoria-context-injector)

• {{FINANCIAL_DISCLOSURE_RULE}} → Jurisdiction-specific rule (e.g., Florida Rule 12.285)
• {{FINANCIAL_DISCLOSURE_DEADLINE}} → Deadline calculation (e.g., 45 days from service)
• {{FINANCIAL_DISCLOSURE_CONSEQUENCES}} → Penalties for non-compliance
• {{FINANCIAL_DISCLOSURE_FORMS}} → Required forms (e.g., Form 12.902(b) or 12.902(c))
• {{CASEMIND_CONTEXT}} → Recent facts from CaseMind
• {{JURISDICTION_RULES}} → Practice knowledge for case jurisdiction

​

Billing Log (victoria-billing-log)

• Tokens used (input/output)
• Model used (Haiku/Sonnet/Opus)
• Cost calculation (based on Anthropic pricing)
• Victoria mode (co_counsel, discovery, financial_analyst, etc.)
• Case ID, firm ID, user ID attribution
• Timestamp and feature type

SELECT
  victoria_mode,
  COUNT(*) as interactions,
  SUM(total_tokens) as tokens,
  SUM(billable_cost) as cost
FROM ai_usage_logs
WHERE firm_id = 'YOUR_FIRM_ID'
  AND created_at > '2025-01-01'
GROUP BY victoria_mode;

​

Cost & Performance Metrics

​

Backend AI Cost Breakdown

• Case Insights: ~$2.40/month ($0.08 × 30 days)
• Memory Extraction: ~$15/month ($0.02 × 25 chats/day × 30 days)
• Task/Deadline Extraction: ~$5/month
• Routing: ~$3/month

​

Performance Metrics

​

Monitoring & Debugging

​

Database Tables

• victoria_case_memory - CaseMind™ extracted facts
• background_extraction_jobs - Queue for all extraction jobs
• case_manager_insights - Cached case insights (24hr TTL)
• ai_usage_logs - Cost tracking and usage metrics
• tasks - Auto-generated and manual tasks
• deadlines - Court dates and filing deadlines

​

SQL Monitoring Queries

SELECT
  job_type,
  priority,
  status,
  COUNT(*) as count
FROM background_extraction_jobs
WHERE created_at > NOW() - INTERVAL '1 hour'
GROUP BY job_type, priority, status
ORDER BY priority DESC, job_type;

SELECT
  c.title as case_title,
  COUNT(m.id) as fact_count,
  MAX(m.created_at) as latest_fact
FROM cases c
LEFT JOIN victoria_case_memory m ON m.case_id = c.id
WHERE c.firm_id = 'YOUR_FIRM_ID'
GROUP BY c.id, c.title
ORDER BY fact_count DESC;

SELECT
  DATE(created_at) as date,
  feature_type,
  SUM(billable_cost) as daily_cost,
  COUNT(*) as executions
FROM ai_usage_logs
WHERE firm_id = 'YOUR_FIRM_ID'
  AND created_at > NOW() - INTERVAL '30 days'
  AND feature_type IN ('casemind_extraction', 'case_insights', 'task_extraction', 'deadline_extraction')
GROUP BY DATE(created_at), feature_type
ORDER BY date DESC, daily_cost DESC;

​

Technical Architecture

​

Event-Driven Design

User Action (Chat/Upload)
       ↓
Main Agent Response (Co-Counsel/Discovery/Financial)
       ↓
Queue Extraction Jobs (via Background Orchestrator)
       ↓
Extraction Worker (Cron job every 2-5 min)
       ↓
Execute Jobs (Memory Extract, Task Extract, etc.)
       ↓
Update Database (victoria_case_memory, tasks, deadlines)
       ↓
Available for Next Agent (via Memory Load, Context Injector)

​

Caching Strategy

• Case Insights: 24-hour cache (expires_at column)
• Practice Knowledge: 5-minute prompt cache (Anthropic)
• Deduplication: 24-hour hash cache (prevents duplicate jobs)
• Memory Facts: No expiration (marked superseded instead)

​

Error Handling

• Retry Logic: 3 attempts with exponential backoff (2s, 4s, 8s)
• Circuit Breakers: Auto-pause on 5 consecutive failures
• Fallback: Graceful degradation (skip extraction if failed)
• Monitoring: All errors logged to ai_usage_logs with status=‘failed’

​

Next Steps