ContextOn.AI OSS

Knowledge Graphs with Confidence Scoring and Failure Learning

By ODEFTO AI Labs | Enterprise Version (ContextOn.AI)

---

What is ContextOn.AI OSS?

ContextOn.AI OSS is an open-source knowledge graph engine for AI agents that introduces confidence-aware, failure-learning graph memory.

Unlike existing tools, ContextOn.AI OSS:

• 🟢 Scores confidence - Every piece of knowledge has a trust indicator
• 🔴 Learns from failures - Marks unreliable knowledge so agents avoid it
• 📊 Shows quality badges - Visual trust indicators (🟢🟡🔴)
• 💡 Suggests questions - Graph tells you what it can answer
• 🔗 Resolves entity aliases - Links "PM-JAY" to "Pradhan Mantri Jan Arogya Yojana"
• 🛠️ Skills - Stores reusable procedures ("how to") with steps and confidence
• 🧰 Tool registry memory - Tracks tools, their descriptions, and which ones fail
• 📦 Auto-context injection - get_context() assembles confident, badge-annotated context for agents
• 🧹 Memory hygiene - Decay sweeps flag stale / low-confidence knowledge for re-verification
• 🎭 Per-agent scoping - Filter knowledge by agent (transparency)
• 🛰️ Works with Claude/Cursor - MCP server included
• 🌐 Web demo - contexton-ai-oss web runs a browser demo of everything

---

Quick Start

Installation

pip install contexton-ai-oss

Basic Usage

from contexton_ai_oss import ContextGraph

# Create a graph
graph = ContextGraph()

# Ingest knowledge from conversations
graph.ingest(
    query="What is PM-JAY?",
    answer="Pradhan Mantri Jan Arogya Yojana is health insurance for poor families",
    agent_id="health-agent"
)

# Query with confidence ranking
results = graph.query("PM-JAY coverage")
for result in results:
    print(f"{result['badge']} {result['node']['content'][:50]}")
    print(f"   Confidence: {result['confidence']:.1%}")

Recording Failures (KEY FEATURE)

# When your agent gives a wrong answer, tell the graph:
graph.record_failure(
    query="What is PM-JAY?",
    answer="It's a housing scheme",
    reason="Incorrect - it's health insurance, not housing"
)

# Confidence in the related knowledge drops (🔴) and failure observations
# are never returned in query results.

# When the agent later gives a correct answer, record the success:
graph.record_success(
    query="What is PM-JAY?",
    answer="It's health insurance for poor families"
)

# Confidence is restored.

Visualizing the Graph

# Generate interactive HTML visualization
graph.visualize("graph.html")
# Open graph.html in your browser

---

Why ContextOn.AI OSS is Different

Feature	Graphify	Graphiti	Mem0	ContextOn.AI OSS
Knowledge graphs	✅	✅	❌	✅
Agent memory	❌	✅	✅	✅
Confidence scoring	❌	❌	❌	✅
Failure learning	❌	❌	❌	✅
Quality badges	❌	❌	❌	✅
Suggested questions	❌	❌	❌	✅
Simple to use	✅	❌	✅	✅

Key Differentiator: No other tool learns from failures. See COMPARISON.md for verified market analysis.

---

Key Features

1. Confidence Scoring

Every node and edge has a confidence score (0.0-1.0) based on:

• How many times it's been verified
• How old the information is
• How many times it's failed

# Get confidence breakdown
node = graph.get_node("node_id")
breakdown = graph.confidence_engine.get_confidence_breakdown(node)
print(breakdown)
# {'mentions': 5, 'base_score': 1.0, 'days_since_verified': 2, ...}

2. Failure Learning (NOVEL)

No other tool learns from mistakes. ContextOn.AI OSS does:

# Record a failure
graph.record_failure(
    query="What is X?",
    answer="Wrong answer",
    reason="Because Y"
)

# The graph now avoids paths that led to this failure
# Future queries prefer more reliable knowledge

3. Quality Badges

See at a glance which knowledge is trustworthy:

• 🟢 High confidence (≥0.8): Verified, reliable
• 🟡 Medium confidence (0.5-0.8): Needs verification
• 🔴 Low confidence (<0.5): Unreliable, verify before using

4. Suggested Questions

The graph analyzes itself and suggests questions it can answer:

suggestions = graph.suggest_questions()
for s in suggestions:
    print(f"{s['badge']} {s['question']}")
    print(f"   Reason: {s['reason']}")

---

Use Cases

1. AI Agent Memory

Give your agents persistent memory that learns and improves:

# Agent remembers past conversations
graph.ingest("How do I reset password?", "Go to settings → security")

# Later, agent can retrieve this knowledge
results = graph.query("password reset")

2. Knowledge Base

Build a knowledge base that tracks reliability:

# Add knowledge with confidence
graph.add_node("PM-JAY covers 5 lakh per family", confidence=0.9)

# Query returns confidence scores
results = graph.query("PM-JAY coverage")

3. Multi-Agent Systems

Agents share knowledge and learn from each other:

# Agent A learns something
graph.ingest("X causes Y", agent_id="agent-a")

# Agent B can query this knowledge
results = graph.query("what causes Y")

---

Integration with AI Assistants

Claude Code / Cursor / Codex (via MCP)

Install with MCP support, then start the server:

pip install "contexton-ai-oss[mcp]"

# stdio transport (recommended for Claude Code / Cursor)
contexton-ai-oss serve

# or streamable HTTP
contexton-ai-oss serve --port 8080

Add to Claude Code's MCP config (.mcp.json):

{
  "mcpServers": {
    "contexton-ai-oss": {
      "command": "contexton-ai-oss",
      "args": ["serve"]
    }
  }
}

Exposed tools: ingest, query, record_failure, record_success, suggest_questions, get_stats, get_aliases, resolve_aliases, get_confidence_breakdown, visualize.

Cursor

# Add to .cursor/rules
ContextOn.AI OSS is available for knowledge retrieval.
Use contexton-ai-oss query for knowledge questions.

Custom Agents

from contexton_ai_oss import ContextGraph

class MyAgent:
    def __init__(self):
        self.memory = ContextGraph()
    
    def answer(self, query):
        # Retrieve relevant knowledge
        knowledge = self.memory.query(query)
        
        # Use knowledge to answer
        return self.generate_answer(query, knowledge)
    
    def learn(self, query, answer, correct):
        if correct:
            self.memory.record_success(query, answer)
        else:
            self.memory.record_failure(query, answer)

---

Command Line Interface

# Ingest a conversation turn
contexton-ai-oss ingest "What is PM-JAY?" "PM-JAY is health insurance for poor families"

# Query with confidence-ranked results
contexton-ai-oss query "PM-JAY coverage"

# Record that an agent gave a wrong / correct answer
contexton-ai-oss record-failure "What is PM-JAY?" "It's a housing scheme" --reason "wrong"
contexton-ai-oss record-success "What is PM-JAY?" "It's health insurance"

# Skills (procedures)
contexton-ai-oss procedure ingest "Reset password" --steps "Open settings; Go to security; Click reset"
contexton-ai-oss procedure get "Reset password"

# Tools
contexton-ai-oss tools register send_email --description "Sends an email"
contexton-ai-oss tools list
contexton-ai-oss tools outcome send_email --error "SMTP timeout"

# Auto-context for agents
contexton-ai-oss context "PM-JAY coverage" --session sess-1

# Memory hygiene + per-agent view
contexton-ai-oss hygiene
contexton-ai-oss agent-memory health-agent

# Statistics, aliases, visualization, browser demo
contexton-ai-oss stats
contexton-ai-oss aliases
contexton-ai-oss visualize graph.html
contexton-ai-oss web --port 8080   # full browser demo

# All commands accept --data-dir DIR to persist the graph to disk

Documentation

• CONCEPTS.md - Why context-aware AI matters
• TUTORIAL.md - Step-by-step guide
• DEPLOYMENT_GUIDE.md - Market comparison, AI deployment fit, hybrid embedding integration
• COMPARISON.md - Verified market comparison
• UPGRADE.md - Enterprise upgrade path
• LAUNCH_STRATEGY.md - Launch plan and checklist

---

Enterprise Features

ContextOn.AI OSS is the open-source version, built by ODEFTO AI Labs. For enterprise features, see ContextOn.AI:

Feature	ContextOn.AI OSS	ContextOn.AI (Enterprise)
Basic graph building	✅	✅
Confidence scoring	✅	✅ Enhanced
Failure learning	✅	✅ Enhanced
Quality badges	✅	✅
Isolation	❌	✅
Quality auditing	❌	✅ 5-dimension scoring
Drift detection	❌	✅
Enterprise connectors	❌	✅ SAP, ServiceNow, Salesforce
Compliance reporting	❌	✅

---

Contributing

We welcome contributions! See CONTRIBUTING.md.

---

License

Apache 2.0 - See LICENSE

---

Support

• 📧 Email: contact@contexton.ai
• 💬 Discussions: GitHub Discussions
• 🐛 Issues: GitHub Issues

---

Acknowledgments

Built by ODEFTO AI Labs.

Inspired by Graphify, Graphiti, and Mem0.