@mytechtoday/augment-extensions 0.2.0 → 0.5.0

package/augment-extensions/domain-rules/mcp/rules/graph-augmented-mcp.md

@@ -0,0 +1,687 @@
+# Graph-Augmented MCP Guidelines
+
+## Overview
+
+**Graph-augmented MCP** uses knowledge graphs to model entities, relationships, and temporal dynamics. This enables structured reasoning, relationship traversal, and context-aware retrieval.
+
+**Key Challenge**: Extracting entities and relationships accurately, maintaining graph consistency, and efficiently traversing large graphs.
+
+---
+
+## 1. Entity Extraction
+
+### Named Entity Recognition (NER)
+
+Extract entities using NER models:
+
+```python
+import spacy
+
+def extract_entities_ner(text: str):
+    """Extract entities using spaCy NER"""
+    nlp = spacy.load("en_core_web_sm")
+    doc = nlp(text)
+    
+    entities = []
+    for ent in doc.ents:
+        entities.append({
+            'text': ent.text,
+            'type': ent.label_,
+            'start': ent.start_char,
+            'end': ent.end_char
+        })
+    
+    return entities
+```
+
+### LLM-Based Extraction
+
+Use LLMs for complex entity extraction:
+
+```python
+def extract_entities_llm(text: str):
+    """Extract entities using LLM"""
+    prompt = f"""
+    Extract all entities from the following text.
+    For each entity, provide:
+    - name: The entity name
+    - type: The entity type (PERSON, ORG, LOCATION, PRODUCT, etc.)
+    - description: Brief description
+    
+    Text: {text}
+    
+    Return as JSON array: [{{"name": "...", "type": "...", "description": "..."}}, ...]
+    """
+    
+    response = llm_call(prompt)
+    entities = json.loads(response)
+    
+    return entities
+```
+
+### Hybrid Extraction
+
+Combine NER and LLM:
+
+```python
+def extract_entities_hybrid(text: str):
+    """Hybrid entity extraction"""
+    # Stage 1: NER for common entities
+    ner_entities = extract_entities_ner(text)
+    
+    # Stage 2: LLM for domain-specific entities
+    llm_entities = extract_entities_llm(text)
+    
+    # Merge and deduplicate
+    all_entities = {}
+    
+    for entity in ner_entities:
+        key = entity['text'].lower()
+        all_entities[key] = entity
+    
+    for entity in llm_entities:
+        key = entity['name'].lower()
+        if key not in all_entities:
+            all_entities[key] = {
+                'text': entity['name'],
+                'type': entity['type'],
+                'description': entity.get('description', '')
+            }
+    
+    return list(all_entities.values())
+```
+
+**Best Practices**:
+- Use NER for speed and common entities
+- Use LLM for domain-specific entities
+- Combine both for best results
+- Validate and deduplicate entities
+
+---
+
+## 2. Relationship Extraction
+
+### Pattern-Based Extraction
+
+Use dependency parsing:
+
+```python
+def extract_relationships_pattern(text: str):
+    """Extract relationships using dependency parsing"""
+    nlp = spacy.load("en_core_web_sm")
+    doc = nlp(text)
+    
+    relationships = []
+    
+    for token in doc:
+        if token.dep_ in ['nsubj', 'dobj']:
+            # Subject-Verb-Object pattern
+            subject = token.text
+            verb = token.head.text
+            
+            # Find object
+            for child in token.head.children:
+                if child.dep_ == 'dobj':
+                    obj = child.text
+                    relationships.append({
+                        'subject': subject,
+                        'predicate': verb,
+                        'object': obj
+                    })
+    
+    return relationships
+```
+
+### LLM-Based Extraction
+
+Use LLMs for complex relationships:
+
+```python
+def extract_relationships_llm(text: str, entities: list):
+    """Extract relationships using LLM"""
+    entity_names = [e['text'] for e in entities]
+    
+    prompt = f"""
+    Given these entities: {', '.join(entity_names)}
+    
+    Extract all relationships from the text.
+    For each relationship, provide:
+    - subject: Entity 1
+    - predicate: Relationship type
+    - object: Entity 2
+    - confidence: 0-1 score
+    
+    Text: {text}
+    
+    Return as JSON array: [{{"subject": "...", "predicate": "...", "object": "...", "confidence": 0.9}}, ...]
+    """
+    
+    response = llm_call(prompt)
+    relationships = json.loads(response)
+    
+    return relationships
+```
+
+**Best Practices**:
+- Extract entities first, then relationships
+- Use pattern-based for simple relationships
+- Use LLM for complex, domain-specific relationships
+- Validate relationship consistency
+
+---
+
+## 3. Graph Modeling
+
+### Graph Schema
+
+Define graph schema:
+
+```python
+from pydantic import BaseModel
+from typing import List, Dict, Optional
+from datetime import datetime
+
+class Entity(BaseModel):
+    id: str
+    name: str
+    type: str
+    properties: Dict[str, any] = {}
+    created_at: datetime
+    updated_at: datetime
+
+class Relationship(BaseModel):
+    id: str
+    subject_id: str
+    predicate: str
+    object_id: str
+    properties: Dict[str, any] = {}
+    confidence: float = 1.0
+    created_at: datetime
+    updated_at: datetime
+
+class KnowledgeGraph(BaseModel):
+    entities: Dict[str, Entity] = {}
+    relationships: List[Relationship] = []
+```
+
+### Graph Construction
+
+Build graph from text:
+
+```python
+class GraphBuilder:
+    """Build knowledge graph from text"""
+    
+    def __init__(self):
+        self.graph = KnowledgeGraph()
+    
+    def add_text(self, text: str):
+        """Add text to graph"""
+        # Extract entities
+        entities = extract_entities_hybrid(text)
+        
+        # Add entities to graph
+        for entity in entities:
+            entity_id = self.generate_id(entity['text'])
+            if entity_id not in self.graph.entities:
+                self.graph.entities[entity_id] = Entity(
+                    id=entity_id,
+                    name=entity['text'],
+                    type=entity['type'],
+                    created_at=datetime.now(),
+                    updated_at=datetime.now()
+                )
+        
+        # Extract relationships
+        relationships = extract_relationships_llm(text, entities)
+        
+        # Add relationships to graph
+        for rel in relationships:
+            subject_id = self.generate_id(rel['subject'])
+            object_id = self.generate_id(rel['object'])
+            
+            if subject_id in self.graph.entities and object_id in self.graph.entities:
+                self.graph.relationships.append(Relationship(
+                    id=self.generate_id(f"{subject_id}-{rel['predicate']}-{object_id}"),
+                    subject_id=subject_id,
+                    predicate=rel['predicate'],
+                    object_id=object_id,
+                    confidence=rel.get('confidence', 1.0),
+                    created_at=datetime.now(),
+                    updated_at=datetime.now()
+                ))
+    
+    def generate_id(self, text: str) -> str:
+        """Generate stable ID from text"""
+        import hashlib
+        return hashlib.md5(text.lower().encode()).hexdigest()[:8]
+```
+
+**Best Practices**:
+- Use stable IDs (hash-based)
+- Version graph schema
+- Validate graph consistency
+- Support incremental updates
+
+---
+
+## 4. Graph Traversal
+
+### Breadth-First Search (BFS)
+
+Find related entities:
+
+```python
+from collections import deque
+
+def bfs_traverse(graph: KnowledgeGraph, start_entity_id: str, max_depth=2):
+    """BFS traversal from start entity"""
+    visited = set()
+    queue = deque([(start_entity_id, 0)])
+    results = []
+
+    while queue:
+        entity_id, depth = queue.popleft()
+
+        if entity_id in visited or depth > max_depth:
+            continue
+
+        visited.add(entity_id)
+
+        # Add entity to results
+        if entity_id in graph.entities:
+            results.append({
+                'entity': graph.entities[entity_id],
+                'depth': depth
+            })
+
+        # Find connected entities
+        for rel in graph.relationships:
+            if rel.subject_id == entity_id and rel.object_id not in visited:
+                queue.append((rel.object_id, depth + 1))
+            elif rel.object_id == entity_id and rel.subject_id not in visited:
+                queue.append((rel.subject_id, depth + 1))
+
+    return results
+```
+
+### Path Finding
+
+Find paths between entities:
+
+```python
+def find_paths(graph: KnowledgeGraph, start_id: str, end_id: str, max_depth=3):
+    """Find all paths between two entities"""
+    paths = []
+
+    def dfs(current_id, target_id, path, visited, depth):
+        if depth > max_depth:
+            return
+
+        if current_id == target_id:
+            paths.append(path.copy())
+            return
+
+        visited.add(current_id)
+
+        # Explore neighbors
+        for rel in graph.relationships:
+            next_id = None
+            if rel.subject_id == current_id and rel.object_id not in visited:
+                next_id = rel.object_id
+            elif rel.object_id == current_id and rel.subject_id not in visited:
+                next_id = rel.subject_id
+
+            if next_id:
+                path.append((rel, next_id))
+                dfs(next_id, target_id, path, visited, depth + 1)
+                path.pop()
+
+        visited.remove(current_id)
+
+    dfs(start_id, end_id, [], set(), 0)
+    return paths
+```
+
+### Subgraph Extraction
+
+Extract relevant subgraph:
+
+```python
+def extract_subgraph(graph: KnowledgeGraph, entity_ids: list, max_depth=1):
+    """Extract subgraph around entities"""
+    subgraph_entities = set(entity_ids)
+    subgraph_relationships = []
+
+    # BFS from each entity
+    for entity_id in entity_ids:
+        neighbors = bfs_traverse(graph, entity_id, max_depth)
+        for neighbor in neighbors:
+            subgraph_entities.add(neighbor['entity'].id)
+
+    # Extract relationships within subgraph
+    for rel in graph.relationships:
+        if rel.subject_id in subgraph_entities and rel.object_id in subgraph_entities:
+            subgraph_relationships.append(rel)
+
+    return KnowledgeGraph(
+        entities={eid: graph.entities[eid] for eid in subgraph_entities if eid in graph.entities},
+        relationships=subgraph_relationships
+    )
+```
+
+**Best Practices**:
+- Limit traversal depth to avoid explosion
+- Use BFS for finding nearby entities
+- Use DFS for path finding
+- Extract subgraphs for context
+
+---
+
+## 5. Retrieval Strategies
+
+### Entity-Based Retrieval
+
+Retrieve by entity:
+
+```python
+def retrieve_by_entity(graph: KnowledgeGraph, entity_name: str, max_depth=2):
+    """Retrieve context around an entity"""
+    # Find entity
+    entity_id = None
+    for eid, entity in graph.entities.items():
+        if entity.name.lower() == entity_name.lower():
+            entity_id = eid
+            break
+
+    if not entity_id:
+        return []
+
+    # Traverse graph
+    results = bfs_traverse(graph, entity_id, max_depth)
+
+    # Format results
+    context = []
+    for result in results:
+        entity = result['entity']
+        depth = result['depth']
+
+        # Find relationships
+        rels = [r for r in graph.relationships
+                if r.subject_id == entity.id or r.object_id == entity.id]
+
+        context.append({
+            'entity': entity.name,
+            'type': entity.type,
+            'depth': depth,
+            'relationships': [
+                f"{graph.entities[r.subject_id].name} {r.predicate} {graph.entities[r.object_id].name}"
+                for r in rels
+                if r.subject_id in graph.entities and r.object_id in graph.entities
+            ]
+        })
+
+    return context
+```
+
+### Relationship-Based Retrieval
+
+Retrieve by relationship type:
+
+```python
+def retrieve_by_relationship(graph: KnowledgeGraph, predicate: str, limit=10):
+    """Retrieve entities connected by relationship type"""
+    results = []
+
+    for rel in graph.relationships:
+        if rel.predicate.lower() == predicate.lower():
+            subject = graph.entities.get(rel.subject_id)
+            obj = graph.entities.get(rel.object_id)
+
+            if subject and obj:
+                results.append({
+                    'subject': subject.name,
+                    'predicate': rel.predicate,
+                    'object': obj.name,
+                    'confidence': rel.confidence
+                })
+
+        if len(results) >= limit:
+            break
+
+    return results
+```
+
+**Best Practices**:
+- Use entity-based retrieval for specific queries
+- Use relationship-based retrieval for pattern queries
+- Combine with vector search for best results
+- Limit depth and breadth to control context size
+
+---
+
+## 6. Temporal Decay
+
+### Time-Based Weighting
+
+Weight entities by recency:
+
+```python
+from datetime import datetime, timedelta
+import math
+
+def apply_temporal_decay(entities: list, decay_rate=0.1):
+    """Apply temporal decay to entity scores"""
+    now = datetime.now()
+
+    for entity in entities:
+        # Calculate age in days
+        age_days = (now - entity['updated_at']).days
+
+        # Apply exponential decay
+        decay_factor = math.exp(-decay_rate * age_days)
+
+        # Update score
+        entity['score'] = entity.get('score', 1.0) * decay_factor
+
+    return sorted(entities, key=lambda x: x['score'], reverse=True)
+```
+
+### Sliding Window
+
+Keep only recent entities:
+
+```python
+def sliding_window_filter(graph: KnowledgeGraph, window_days=30):
+    """Filter graph to recent entities"""
+    cutoff = datetime.now() - timedelta(days=window_days)
+
+    # Filter entities
+    recent_entities = {
+        eid: entity
+        for eid, entity in graph.entities.items()
+        if entity.updated_at >= cutoff
+    }
+
+    # Filter relationships
+    recent_relationships = [
+        rel for rel in graph.relationships
+        if rel.updated_at >= cutoff
+        and rel.subject_id in recent_entities
+        and rel.object_id in recent_entities
+    ]
+
+    return KnowledgeGraph(
+        entities=recent_entities,
+        relationships=recent_relationships
+    )
+```
+
+**Best Practices**:
+- Apply temporal decay for long-running agents
+- Use sliding windows to limit graph size
+- Adjust decay rate based on domain
+- Preserve important entities regardless of age
+
+---
+
+## 7. Graph Database Options
+
+### Comparison
+
+| Database | Type | Scale | Query Language | Features |
+|----------|------|-------|----------------|----------|
+| **Neo4j** | Native graph | Billions | Cypher | ACID, clustering, visualization |
+| **Amazon Neptune** | Managed | Billions | Gremlin/SPARQL | Serverless, high availability |
+| **ArangoDB** | Multi-model | Millions | AQL | Document + Graph, flexible |
+| **NetworkX** | In-memory | Thousands | Python API | Simple, local, analysis |
+
+### Neo4j Example
+
+```python
+from neo4j import GraphDatabase
+
+class Neo4jGraph:
+    def __init__(self, uri, user, password):
+        self.driver = GraphDatabase.driver(uri, auth=(user, password))
+
+    def add_entity(self, entity: Entity):
+        """Add entity to Neo4j"""
+        with self.driver.session() as session:
+            session.run(
+                "MERGE (e:Entity {id: $id}) "
+                "SET e.name = $name, e.type = $type, e.updated_at = datetime()",
+                id=entity.id, name=entity.name, type=entity.type
+            )
+
+    def add_relationship(self, rel: Relationship):
+        """Add relationship to Neo4j"""
+        with self.driver.session() as session:
+            session.run(
+                "MATCH (a:Entity {id: $subject_id}), (b:Entity {id: $object_id}) "
+                "MERGE (a)-[r:RELATES {predicate: $predicate}]->(b) "
+                "SET r.confidence = $confidence, r.updated_at = datetime()",
+                subject_id=rel.subject_id,
+                object_id=rel.object_id,
+                predicate=rel.predicate,
+                confidence=rel.confidence
+            )
+
+    def traverse(self, entity_id: str, max_depth=2):
+        """Traverse graph from entity"""
+        with self.driver.session() as session:
+            result = session.run(
+                "MATCH path = (start:Entity {id: $id})-[*1..$depth]-(end:Entity) "
+                "RETURN path",
+                id=entity_id, depth=max_depth
+            )
+            return [record["path"] for record in result]
+```
+
+**Best Practices**:
+- Use Neo4j for production graph workloads
+- Use NetworkX for prototyping and analysis
+- Use managed services (Neptune) for scale
+- Optimize queries with indexes
+
+---
+
+## 8. Best Practices
+
+### DO
+
+✅ **Extract entities accurately**: Use hybrid NER + LLM
+✅ **Validate relationships**: Ensure consistency
+✅ **Limit traversal depth**: Avoid graph explosion
+✅ **Apply temporal decay**: Keep graph relevant
+✅ **Use graph databases**: For large-scale graphs
+✅ **Combine with vector search**: Best of both worlds
+✅ **Version graph schema**: Support evolution
+✅ **Monitor graph size**: Prune periodically
+
+### DON'T
+
+❌ **Don't extract all entities**: Be selective
+❌ **Don't ignore confidence scores**: Use for filtering
+❌ **Don't traverse entire graph**: Limit scope
+❌ **Don't keep stale entities**: Apply decay
+❌ **Don't use in-memory for large graphs**: Use database
+❌ **Don't forget to deduplicate**: Entities and relationships
+❌ **Don't ignore graph structure**: Leverage relationships
+
+---
+
+## 9. Common Pitfalls
+
+### Graph Explosion
+
+**Problem**: Graph grows too large, traversal too slow
+
+**Solution**:
+- Limit traversal depth (1-3 hops)
+- Apply temporal decay
+- Prune low-confidence relationships
+- Use subgraph extraction
+
+### Poor Entity Extraction
+
+**Problem**: Missing or incorrect entities
+
+**Solution**:
+- Use hybrid NER + LLM extraction
+- Validate entities against schema
+- Implement entity resolution (deduplication)
+- Use domain-specific models
+
+### Relationship Noise
+
+**Problem**: Too many low-quality relationships
+
+**Solution**:
+- Filter by confidence threshold
+- Validate relationship types
+- Use pattern-based extraction for precision
+- Review and prune periodically
+
+---
+
+## Configuration Example
+
+```json
+{
+  "mcp": {
+    "type": "graph",
+    "entityExtraction": {
+      "method": "hybrid",
+      "nerModel": "en_core_web_sm",
+      "llmModel": "gpt-4o-mini",
+      "confidenceThreshold": 0.7
+    },
+    "relationshipExtraction": {
+      "method": "llm",
+      "llmModel": "gpt-4o-mini",
+      "confidenceThreshold": 0.8
+    },
+    "graphDatabase": {
+      "provider": "neo4j",
+      "uri": "bolt://localhost:7687",
+      "maxConnections": 50
+    },
+    "traversal": {
+      "maxDepth": 2,
+      "maxResults": 50
+    },
+    "temporalDecay": {
+      "enabled": true,
+      "decayRate": 0.1,
+      "slidingWindowDays": 30
+    },
+    "tokenBudget": {
+      "maxGraphTokens": 2048
+    }
+  }
+}
+```
+
+---
+