Knowledge Graph

Knowledge Graph Specialist

Overview

knowledge_graph wraps patterns from abhigyanpatwari/GitNexus for code entity extraction and graph construction, and topoteretes/cognee for Graph RAG retrieval.

Given any source string, the specialist hashes that string and constructs a synthetic graph-shaped response. It returns artificial entity lists, relevance scores, and relationship paths for exercising downstream contracts. It does not parse or cite the named artifact.

The specialist is stateless. Each execute call creates a new graph UUID and self-contained result; there is no previously built graph to retrieve across calls.

Those descriptions are interface goals, not current data processing. The implementation hashes the source string and generates artificial counts, entities, scores, and paths; it has no stored graph.

Prerequisites

• Use Python 3.11+ in a local OSS Agent Lab checkout and run pip install -e ..
• Pass only a source identifier or sample text; paths and URLs are not opened.
• Read the runtime contract before using the response.

Capabilities

• knowledge_graph: Generate graph-shaped counts and type labels from a source-string hash.
• code_analysis: Exercise a code-analysis response schema with artificial entities.
• entity_linking: Generate deterministic-style IDs and synthetic paths for contract tests.
• graph_rag: Exercise ranking and path fields without retrieval or generation.

Tools

Parameters

Request-level (intent.parameters)

Graph types

• code — labels output with module, class, function, variable, and import types.
• document — labels output with concept, section, claim, and reference types.
• mixed — returns a combined synthetic type list.

Output

{
    "graph": {
        "graph_id": str,          # stable UUID for this graph instance
        "node_count": int,        # total entity nodes
        "edge_count": int,        # total directed relationship edges
        "entity_types": list[str],# entity labels present in the graph
        "graph_type": str,        # echo of requested graph_type
    },
    "query_results": {
        "results": list[dict],    # ranked entity hits
        "relevance_scores": list[float],  # parallel relevance values 0-1
        "paths": list[dict],      # connecting relationship paths
        "total_found": int,       # total matches before max_results cap
    },
    # only present when entity_a and entity_b are supplied:
    "relationships": {
        "paths": list[dict],      # all paths from entity_a to entity_b
        "relationship_types": list[str],  # deduplicated edge labels
        "strength": float,        # coupling strength 0-1
        "path_count": int,
    },
}

Instructions

1. Supply a non-empty source string and query plus a supported graph type and bounded positive limits.
2. Run the specialist through the Python API or CLI.
3. Treat graph IDs as per-call identifiers and every entity/path as artificial fixture data.
4. Use a real parser or graph store when the task requires claims about source code.

Examples

Python API

import asyncio
from agents.specialists.knowledge_graph.agent import KnowledgeGraphSpecialist
from oss_agent_lab.contracts import Intent, Query, SpecialistRequest

specialist = KnowledgeGraphSpecialist()

request = SpecialistRequest(
    intent=Intent(
        action="knowledge_graph",
        domain="code_analysis",
        confidence=0.9,
        parameters={
            "source": "https://github.com/abhigyanpatwari/GitNexus",
            "graph_type": "code",
            "max_depth": 3,
        },
    ),
    query=Query(user_input="Which functions call the authentication module?"),
    specialist_name="knowledge_graph",
)

response = asyncio.run(specialist.execute(request))
print(response.result["graph"]["node_count"])
print(response.result["query_results"]["results"][0])

CLI

oss-lab run knowledge_graph "Which functions call the authentication module?"

With relationship search

request = SpecialistRequest(
    intent=Intent(
        action="knowledge_graph",
        domain="code_analysis",
        confidence=0.9,
        parameters={
            "source": "./src/",
            "entity_a": "UserService",
            "entity_b": "DatabaseAdapter",
        },
    ),
    query=Query(user_input="How does UserService depend on DatabaseAdapter?"),
    specialist_name="knowledge_graph",
)

CLI with parameters

oss-lab run knowledge_graph "dependency chain for PaymentProcessor" \
  --param source=https://github.com/owner/repo \
  --param graph_type=code \
  --param max_depth=5 \
  --param entity_a=PaymentProcessor \
  --param entity_b=DatabaseClient

Error Handling

• Reject empty source or query values, unknown graph types, and non-positive depth/result limits.
• Do not claim that a path or URL was read merely because the result contains a graph ID.
• Do not cite synthetic entities or relationships as evidence about a repository.

Resources

Wraps abhigyanpatwari/GitNexus and topoteretes/cognee.

The current local implementation is interface-only. See the runtime contract.