Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Latest commit

 

History

History

graph_overlay_examples

README.md

SochDB New Features Examples

This directory contains examples demonstrating graph-related workflows in the current SDK. Some previously documented features (GraphOverlay, Policy Hooks, Tool Routing) are not exposed in the Python SDK yet, so those examples show app-level equivalents.

Features Demonstrated

1. Graph API (graph_overlay_example.py)

Lightweight graph layer on KV storage for agent memory:

  • Node Operations: Create, read, update, delete nodes with types and properties
  • Edge Operations: Typed edges with bidirectional indexes
  • Traversal: BFS and DFS traversal algorithms
  • Queries: Find neighbors, get nodes by type, shortest path
from sochdb import GraphOverlay, GraphNode, GraphEdge

db = sochdb.open("./db")
graph = GraphOverlay(db, namespace="memory")

# Create nodes
graph.add_node(GraphNode(id="user_1", type="person", properties={"name": "Alice"}))

# Create edges
graph.add_edge(GraphEdge(from_id="user_1", edge_type="knows", to_id="user_2"))

# Traverse
visited = graph.bfs("user_1", max_depth=3)

2. Policy Hooks (App-Level) (policy_hooks_example.py)

Trigger-based guardrails for agent operations:

  • Validation Hooks: @before_write for input validation
  • Redaction Hooks: @after_read for PII masking
  • Access Control: @before_delete for permission checks
  • Audit Logging: @after_commit for compliance
  • Rate Limiting: Token bucket algorithm per agent/session
from sochdb.policy import PolicyEngine, before_write, PolicyAction

engine = PolicyEngine(db)

@before_write(engine, pattern="users/*")
def validate_user(ctx):
    if "email" not in ctx.value:
        return PolicyAction.DENY
    return PolicyAction.ALLOW

3. Tool Routing (App-Level) (tool_routing_example.py)

Context-driven dynamic binding for multi-agent systems:

  • Agent Registry: Register agents with capabilities
  • Routing Strategies: PRIORITY, ROUND_ROBIN, LEAST_LOADED, STICKY, RANDOM, FASTEST
  • Tool Categories: CODE, SEARCH, DATABASE, MEMORY, VECTOR, GRAPH, ANALYTICS, API, FILE, SYSTEM
  • Failover: Automatic failover to backup agents
from sochdb.routing import AgentRegistry, ToolRouter, ToolCategory, RoutingStrategy

registry = AgentRegistry(db)
router = ToolRouter(registry, default_strategy=RoutingStrategy.PRIORITY)

# Register agent
registry.register(
    agent_id="code_analyzer",
    capabilities=[ToolCategory.CODE, ToolCategory.SEARCH],
    priority=100
)

# Route tool call
result = router.route(RouteContext(tool_name="search_code"))

4. Context Query (context_query_example.py)

Token-aware retrieval for LLM applications:

  • Token Budgeting: Fit context within LLM token limits
  • Deduplication: Semantic and exact deduplication
  • Hybrid Search: Vector + keyword with RRF fusion
  • Provenance: Track source of each chunk
  • Token Estimators: GPT-4, Claude, or custom
from sochdb.context import ContextQuery, DeduplicationStrategy

query = (ContextQuery(db, collection="docs")
    .with_token_budget(4000)
    .with_query("How does vector search work?")
    .with_deduplication(DeduplicationStrategy.SEMANTIC)
    .with_top_k(10))

result = query.execute()
print(f"Used {result.total_tokens} tokens")

Running the Examples

  1. Install the SochDB Python SDK (venv in this repo):

    ./../venv/bin/pip install sochdb

  2. Run any example:

    ./../venv/bin/python graph_overlay_example.py
./../venv/bin/python policy_hooks_example.py
./../venv/bin/python tool_routing_example.py
./../venv/bin/python context_query_example.py

Related Documentation

Requirements

  • Python 3.8+
  • SochDB v0.3.x or later