Autodidact
Autonomous Self-Learning

How It Works

Autodidact runs a continuous loop: detect gaps in your knowledge graph, generate research questions, search evidence sources, extract structured claims via LLM, and ingest validated findings. Each cycle closes blindspots and raises confidence on under-evidenced entities.

Quick Start

from attestdb import AttestDB

db = AttestDB("my.attest")

# Enable with defaults: 1-hour cycle, $1/day cap, 100 LLM calls/day
status = db.enable_autodidact()
print(f"Next cycle at: {status.next_cycle_at}")

# Check what it will cost
estimate = db.autodidact_cost_estimate(cycles=24)
print(f"Estimated daily cost: ${estimate['daily']:.2f}")

# Check current status
status = db.autodidact_status()
print(f"Cycles completed: {status.cycle_count}")
print(f"Claims ingested: {status.total_claims_ingested}")
print(f"Cost today: ${status.estimated_cost_today:.2f}")

# Review history
for report in db.autodidact_history(limit=5):
    print(f"Cycle {report.cycle_number}: {report.claims_ingested} claims, "
          f"{report.negative_results} negatives, ${report.estimated_cost:.3f}")

# Trigger an immediate cycle (doesn't wait for timer)
db.autodidact_run_now()

# Disable when done
db.disable_autodidact()

Configuration

enable_autodidact() accepts fine-grained controls for budget, scope, and behavior:

Evidence Sources

Autodidact can search multiple evidence providers. Free sources require no API key.

# Use free sources only (no API key needed)
db.enable_autodidact(sources="pubmed")

# Use multiple sources
db.enable_autodidact(sources="auto")  # auto-detects available API keys

# Custom search function
def my_search(query):
    return [{"title": "...", "text": "...", "url": "..."}]

db.enable_autodidact(search_fn=my_search)

Budget Controls

Autodidact will never exceed your configured cost cap. When the daily budget is exhausted, the daemon pauses until the next day. Every cycle reports its cost.

# Conservative: $0.50/day, 50 LLM calls, 30-minute cycles
db.enable_autodidact(
    interval=1800,
    max_cost_per_day=0.50,
    max_llm_calls_per_day=50,
    max_questions_per_cycle=3,
)

# Check cost estimate before enabling
estimate = db.autodidact_cost_estimate(cycles=24)
print(f"Per cycle: ${estimate['per_cycle']:.3f}")
print(f"Daily (24 cycles): ${estimate['daily']:.2f}")
print(f"Monthly: ${estimate['monthly']:.2f}")

Event Triggers

Beyond the timer-based cycle, autodidact can react to events in the knowledge graph:

# Enable all triggers
db.enable_autodidact(
    enabled_triggers=["timer", "retraction", "inquiry"],
    trigger_cooldown=60.0,  # min seconds between event-triggered cycles
)

# Register a question — triggers immediate research if inquiry trigger is enabled
db.register_inquiry("What are the off-target effects of CRISPR-Cas9 in hepatocytes?")

Cycle Reports

Every cycle produces a CycleReport with detailed metrics:

REST API

The enterprise API exposes autodidact via 6 endpoints. All require an API key with admin role.

# Enable via REST API
curl -X POST https://api.attestdb.com/api/v1/autodidact/enable \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"interval": 3600, "max_cost_per_day": 1.0}'

# Check status
curl https://api.attestdb.com/api/v1/autodidact/status \
  -H 'Authorization: Bearer YOUR_API_KEY'

# Get cost estimate for 24 cycles
curl 'https://api.attestdb.com/api/v1/autodidact/cost-estimate?cycles=24' \
  -H 'Authorization: Bearer YOUR_API_KEY'

MCP Tools

Autodidact is also accessible via 5 MCP tools for AI agent integration: