allycat / 2b_process_graph_phase3.py
niloydebbarma's picture
Upload 50 files
a7d2416 verified
"""
Phase 3: Community Summarization using LLM
Loads graph-data-phase-2.json, generates summaries, saves graph-data-final.json
"""
import json
import logging
import os
import time
from pathlib import Path
from typing import Dict, Any, List
from datetime import datetime
from collections import defaultdict
import networkx as nx
import openai
import google.generativeai as genai
# JSON parsing libraries (same as Phase 1)
import orjson
from json_repair import repair_json
from my_config import MY_CONFIG
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
class GraphBuilderPhase3:
"""Phase 3: Generate community summaries using LLM"""
def __init__(self, llm_provider: str = "cerebras"):
"""Initialize Phase 3 processor"""
self.llm_provider = llm_provider.lower()
self.graph_data = None
self.nx_graph = None
self.community_assignments = {}
self.community_stats = {}
# Initialize LLM API based on provider
if self.llm_provider == "cerebras":
if not MY_CONFIG.CEREBRAS_API_KEY:
raise ValueError("CEREBRAS_API_KEY not set")
self.cerebras_client = openai.OpenAI(
api_key=MY_CONFIG.CEREBRAS_API_KEY,
base_url="https://api.cerebras.ai/v1"
)
self.model_name = "llama-4-scout-17b-16e-instruct"
logger.info("πŸš€ Using Cerebras API")
elif self.llm_provider == "gemini":
if not MY_CONFIG.GEMINI_API_KEY:
raise ValueError("GEMINI_API_KEY not set")
genai.configure(api_key=MY_CONFIG.GEMINI_API_KEY)
self.model_name = "gemini-1.5-flash"
self.gemini_model = genai.GenerativeModel(self.model_name)
logger.info("πŸ†“ Using Google Gemini API")
else:
raise ValueError(f"Invalid provider '{llm_provider}'. Choose: cerebras, gemini")
# Initialize embedding model for DRIFT search metadata
try:
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
self.embedding_model = HuggingFaceEmbedding(
model_name=MY_CONFIG.EMBEDDING_MODEL
)
logger.info(f"πŸ” Initialized embedding model: {MY_CONFIG.EMBEDDING_MODEL}")
except Exception as e:
logger.warning(f"⚠️ Embedding model initialization failed: {e}")
self.embedding_model = None
logger.info("βœ… Phase 3 initialized: Community Summarization")
logger.info(f"πŸ“Š LLM Provider: {self.llm_provider.upper()}, Model: {self.model_name}")
# STEP 1: Load Phase 2 Output
def load_graph_data(self, input_path: str = None) -> bool:
"""Load graph-data-phase-2.json from Phase 2"""
if input_path is None:
input_path = "workspace/graph_data/graph-data-phase-2.json"
try:
input_file = Path(input_path)
if not input_file.exists():
logger.error(f"❌ Input file not found: {input_path}")
logger.error(" Please run Phase 2 (2b_process_graph_phase2.py) first")
return False
with open(input_file, 'r', encoding='utf-8') as f:
self.graph_data = json.load(f)
node_count = len(self.graph_data.get("nodes", []))
rel_count = len(self.graph_data.get("relationships", []))
# Verify Phase 2 was completed
if self.graph_data.get("metadata", {}).get("phase") != "community_detection":
logger.error("❌ Input file is not from Phase 2 (community_detection)")
return False
logger.info(f"πŸ“‚ Loaded graph-data-phase-2.json: {node_count} nodes, {rel_count} relationships")
# Load community stats
self.community_stats = self.graph_data.get("community_stats", {})
num_communities = len(self.community_stats)
logger.info(f"πŸ“Š Found {num_communities} communities to summarize")
if num_communities == 0:
logger.error("❌ No communities found in Phase 2 output")
return False
return True
except Exception as e:
logger.error(f"❌ Error loading graph data: {e}")
return False
# STEP 2: Build NetworkX Graph
def _build_networkx_graph(self) -> nx.Graph:
"""Rebuild NetworkX graph from JSON data"""
logger.info("πŸ”¨ Building NetworkX graph from JSON data...")
G = nx.Graph()
# Add nodes with attributes
for node in self.graph_data["nodes"]:
node_id = node["id"]
properties = node.get("properties", {})
G.add_node(
node_id,
name=properties.get("name", ""),
type=node.get("labels", ["Unknown"])[0],
description=properties.get("content", ""),
community_id=properties.get("community_id", ""),
degree_centrality=properties.get("degree_centrality", 0.0)
)
# Add edges
for rel in self.graph_data["relationships"]:
start_node = rel.get("startNode")
end_node = rel.get("endNode")
if start_node in G.nodes() and end_node in G.nodes():
G.add_edge(start_node, end_node)
logger.info(f"βœ… Built NetworkX graph: {G.number_of_nodes()} nodes, {G.number_of_edges()} edges")
return G
# STEP 3: Extract Community Assignments
def _extract_community_assignments(self) -> Dict[str, int]:
"""Extract community assignments from node properties"""
logger.info("πŸ“‹ Extracting community assignments from nodes...")
assignments = {}
for node in self.graph_data["nodes"]:
node_id = node["id"]
comm_id_str = node.get("properties", {}).get("community_id", "")
if comm_id_str and comm_id_str.startswith("comm-"):
try:
comm_id = int(comm_id_str.replace("comm-", ""))
assignments[node_id] = comm_id
except ValueError:
logger.warning(f"Invalid community_id format: {comm_id_str}")
logger.info(f"βœ… Extracted {len(assignments)} community assignments")
return assignments
# STEP 4: LLM Inference Methods
def _cerebras_inference(self, system_prompt: str, user_prompt: str) -> str:
"""Call Cerebras API for inference"""
try:
# Calculate dynamic parameters based on community size and complexity
total_nodes = self.nx_graph.number_of_nodes() if hasattr(self, 'nx_graph') else 100
complexity_factor = min(1.0, total_nodes / 1000)
# Adaptive temperature: higher for complex graphs to encourage creativity
dynamic_temperature = round(0.1 + (complexity_factor * 0.4), 2) # Range: 0.1-0.5
# Adaptive tokens: more for larger/complex summaries
dynamic_tokens = int(300 + (complexity_factor * 400)) # Range: 300-700
response = self.cerebras_client.chat.completions.create(
model=self.model_name,
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt}
],
temperature=dynamic_temperature,
max_tokens=dynamic_tokens
)
if not response or not response.choices or not response.choices[0].message.content:
raise ValueError("Empty response from Cerebras")
return response.choices[0].message.content.strip()
except Exception as e:
logger.error(f"Cerebras inference error: {e}")
raise e
def _gemini_inference(self, system_prompt: str, user_prompt: str) -> str:
"""Call Gemini API for inference"""
try:
# Calculate dynamic generation config based on graph complexity
total_nodes = self.nx_graph.number_of_nodes() if hasattr(self, 'nx_graph') else 100
complexity_factor = min(1.0, total_nodes / 1000)
# Adaptive temperature and tokens for Gemini
dynamic_temperature = round(0.1 + (complexity_factor * 0.4), 2)
dynamic_tokens = int(300 + (complexity_factor * 400))
generation_config = {
"temperature": dynamic_temperature,
"max_output_tokens": dynamic_tokens,
"candidate_count": 1
}
combined_prompt = f"{system_prompt}\n\n{user_prompt}"
response = self.gemini_model.generate_content(
combined_prompt,
generation_config=generation_config
)
if not response or not response.text:
raise ValueError("Empty response from Gemini")
return response.text.strip()
except Exception as e:
logger.error(f"Gemini inference error: {e}")
raise e
# STEP 5: Generate Community Summaries
def _generate_community_summaries(self) -> Dict[int, str]:
"""Generate LLM summaries for each community"""
logger.info("πŸ“ Generating community summaries with LLM...")
logger.info(f" Total communities to summarize: {len(self.community_stats)}")
summaries = {}
# Group nodes by community
communities = defaultdict(list)
for node_id, comm_id in self.community_assignments.items():
communities[comm_id].append(node_id)
start_time = time.time()
for idx, (comm_id_str, stats) in enumerate(self.community_stats.items(), 1):
comm_id = int(comm_id_str)
logger.info(f" Processing community {idx}/{len(self.community_stats)}: comm-{comm_id} ({stats['member_count']} members)")
# Get top entities by centrality
node_ids = communities[comm_id]
subgraph = self.nx_graph.subgraph(node_ids)
# Get nodes sorted by degree centrality
centrality = nx.degree_centrality(subgraph)
top_nodes = sorted(centrality.items(), key=lambda x: x[1], reverse=True)[:15]
# Prepare entity information for LLM
entity_info = []
for node_id, _ in top_nodes:
node_data = self.nx_graph.nodes[node_id]
entity_info.append({
"name": node_data.get("name", "Unknown"),
"type": node_data.get("type", "Unknown"),
"description": node_data.get("description", "")[:150] # Limit length
})
# Create LLM prompt
# Senior-developer style system/user prompts with strict output schema
# Calculate dynamic topic count based on community size
topic_count = max(2, min(5, stats['member_count'] // 3)) # Scale with community size
system_prompt = (
"You are a specialized knowledge graph summarization assistant. Your task is to analyze community "
"structures and generate comprehensive summaries for graph-based retrieval systems.\n\n"
"CONSTITUTIONAL AI PRINCIPLES:\n"
"1. Content-Adaptive: Generate summaries based on actual community composition and statistics\n"
"2. Context-Aware: Consider entity relationships and community density in summarization\n"
"3. Quality-First: Prioritize accuracy and relevance over brevity\n"
"4. Structured Output: Ensure consistent JSON format for programmatic consumption\n\n"
"SUMMARIZATION GUIDELINES:\n"
"- Analyze entity types, relationships, and community structure\n"
"- Identify key themes and concepts that define this community\n"
"- Generate topics that capture semantic meaning, not just entity names\n"
"- Assess confidence based on data completeness and coherence\n"
"- Use neutral, factual tone suitable for technical documentation"
)
user_prompt = (
f"Analyze the following community data and generate a structured summary.\n\n"
f"COMMUNITY STATISTICS:\n"
f"- Total Members: {stats['member_count']}\n"
f"- Internal Connections: {stats['internal_edges']}\n"
f"- Community Density: {stats['density']:.3f}\n"
f"- Connectivity Strength: {'High' if stats['density'] > 0.1 else 'Medium' if stats['density'] > 0.05 else 'Low'}\n\n"
f"TOP ENTITIES (name, type, description):\n{json.dumps(entity_info, indent=2)}\n\n"
f"OUTPUT FORMAT (strict JSON):\n"
f"{{\n"
f" \"summary\": \"2-3 sentence comprehensive summary of community purpose and characteristics\",\n"
f" \"primary_topics\": [\"topic_1\", \"topic_2\", \"topic_{topic_count}\"],\n"
f" \"confidence\": 0.85\n"
f"}}\n\n"
f"VALIDATION REQUIREMENTS:\n"
f"- summary: Must be 2-3 complete sentences describing community focus and key characteristics\n"
f"- primary_topics: Array of exactly {topic_count} descriptive phrases (not just entity names)\n"
f"- confidence: Float between 0.0-1.0 based on data quality and coherence\n\n"
f"IMPORTANT: Respond with ONLY the JSON object. No markdown formatting, no explanations, no code blocks."
)
# Call LLM for summary
try:
if self.llm_provider == "gemini":
summary_response = self._gemini_inference(system_prompt, user_prompt)
else: # cerebras
summary_response = self._cerebras_inference(system_prompt, user_prompt)
# Parse JSON response
parsed_summary = self._parse_summary_response(summary_response, comm_id)
if parsed_summary:
summaries[comm_id] = parsed_summary
else:
# Fallback to raw response if parsing fails
summaries[comm_id] = summary_response.strip()
# Log progress every 10 communities
if idx % 10 == 0:
elapsed = time.time() - start_time
avg_time = elapsed / idx
remaining = avg_time * (len(self.community_stats) - idx)
logger.info(f" Progress: {idx}/{len(self.community_stats)} ({elapsed:.1f}s elapsed, ~{remaining:.1f}s remaining)")
except Exception as e:
logger.error(f"❌ Failed to generate summary for community {comm_id}: {e}")
summaries[comm_id] = f"Community with {stats['member_count']} entities focused on {entity_info[0]['type'] if entity_info else 'various'} topics."
elapsed = time.time() - start_time
logger.info(f"βœ… Generated {len(summaries)} community summaries in {elapsed:.1f}s")
return summaries
def _parse_summary_response(self, response: str, comm_id: int) -> str:
"""Parse JSON summary response with fallback to text extraction"""
try:
# Clean response
cleaned_response = response.strip()
# Remove markdown formatting
if "```json" in cleaned_response:
parts = cleaned_response.split("```json")
if len(parts) > 1:
json_part = parts[1].split("```")[0].strip()
cleaned_response = json_part
elif "```" in cleaned_response:
parts = cleaned_response.split("```")
if len(parts) >= 3:
cleaned_response = parts[1].strip()
# Try to parse JSON
try:
summary_data = self._smart_json_parse_summary(cleaned_response)
if summary_data and isinstance(summary_data, dict):
summary_text = summary_data.get('summary', '')
if summary_text and len(summary_text.strip()) > 10:
return summary_text.strip()
except ValueError as e:
logger.debug(f"Summary JSON parsing failed for comm-{comm_id}: {e}")
except Exception as e:
logger.debug(f"Summary JSON parsing unexpected error for comm-{comm_id}: {e}")
except Exception as e:
logger.debug(f"Summary JSON parsing failed for comm-{comm_id}: {e}")
# Fallback: extract first meaningful sentence
try:
lines = response.split('\n')
for line in lines:
line = line.strip()
if len(line) > 20 and '.' in line and not line.startswith('{'):
return line
except Exception:
pass
return None
def _smart_json_parse_summary(self, json_text: str) -> Dict:
"""
Simple 5-step JSON parsing approach (exactly same as Phase 1)
"""
cleaned_text = json_text.strip()
# Step 1: orjson
try:
result = orjson.loads(cleaned_text.encode('utf-8'))
logger.debug("βœ… Step 1: orjson succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 1: orjson failed - {e}")
# Step 2: json-repair
try:
repaired = repair_json(cleaned_text)
result = orjson.loads(repaired.encode('utf-8'))
logger.debug("βœ… Step 2: json-repair + orjson succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 2: json-repair failed - {e}")
# Step 3: standard json
try:
result = json.loads(cleaned_text)
logger.debug("βœ… Step 3: standard json succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 3: standard json failed - {e}")
# Step 4: json-repair + standard json
try:
repaired = repair_json(cleaned_text)
result = json.loads(repaired)
logger.debug("βœ… Step 4: json-repair + standard json succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 4: json-repair + standard json failed - {e}")
# Step 5: All failed - this will trigger save failed txt files
raise ValueError("All 4 JSON parsing steps failed")
# STEP 6: Identify Key Entities
def _identify_key_entities(self) -> Dict[int, List[str]]:
"""Identify key entities in each community based on centrality"""
logger.info("πŸ”‘ Identifying key entities per community...")
key_entities = {}
# Group nodes by community
communities = defaultdict(list)
for node_id, comm_id in self.community_assignments.items():
communities[comm_id].append(node_id)
for comm_id, node_ids in communities.items():
subgraph = self.nx_graph.subgraph(node_ids)
# Calculate degree centrality
centrality = nx.degree_centrality(subgraph)
# Get top 5 entities
top_nodes = sorted(centrality.items(), key=lambda x: x[1], reverse=True)[:5]
key_entities[comm_id] = [
self.nx_graph.nodes[node_id].get("name", "Unknown")
for node_id, _ in top_nodes
]
logger.info(f"βœ… Identified key entities for {len(key_entities)} communities")
return key_entities
# STEP 7: Create Community Nodes
def _create_community_nodes(self, community_summaries: Dict[int, str], key_entities: Dict[int, List[str]]) -> List[Dict]:
"""Create community nodes for the graph"""
logger.info("πŸ—οΈ Creating community nodes...")
import uuid
community_nodes = []
for comm_id_str, stats in self.community_stats.items():
comm_id = int(comm_id_str)
node = {
"id": f"community-{uuid.uuid4()}",
"elementId": f"community-{uuid.uuid4()}",
"labels": ["Community"],
"properties": {
"community_id": f"comm-{comm_id}",
"level": 1,
"member_count": stats["member_count"],
"internal_edges": stats["internal_edges"],
"density": round(stats["density"], 4),
"avg_degree": round(stats["avg_degree"], 2),
"summary": community_summaries.get(comm_id, ""),
"key_entities": key_entities.get(comm_id, []),
"created_date": datetime.now().isoformat()
}
}
community_nodes.append(node)
logger.info(f"βœ… Created {len(community_nodes)} community nodes")
return community_nodes
# STEP 8: Create IN_COMMUNITY Relationships
def _create_in_community_relationships(self, community_nodes: List[Dict]) -> List[Dict]:
"""Create IN_COMMUNITY relationships linking entities to communities"""
logger.info("Creating IN_COMMUNITY relationships...")
import uuid
# Create mapping from community_id to community node id
comm_id_to_node_id = {}
for node in community_nodes:
comm_id = node["properties"]["community_id"]
comm_id_to_node_id[comm_id] = node["id"]
relationships = []
for entity_id, comm_id in self.community_assignments.items():
comm_node_id = comm_id_to_node_id.get(f"comm-{comm_id}")
if comm_node_id:
# Calculate confidence based on community membership strength
entity_node = next((n for n in self.graph_data['nodes'] if n['id'] == entity_id), None)
if entity_node:
degree_centrality = entity_node.get('properties', {}).get('degree_centrality', 0.5)
# Higher centrality = higher confidence in community assignment
dynamic_confidence = round(0.6 + (degree_centrality * 0.4), 3) # Range: 0.6-1.0
else:
dynamic_confidence = 0.8 # Default for missing nodes
rel = {
"id": f"rel-{uuid.uuid4()}",
"startNode": entity_id,
"endNode": comm_node_id,
"type": "IN_COMMUNITY",
"properties": {
"confidence": dynamic_confidence,
"assigned_date": datetime.now().isoformat()
}
}
relationships.append(rel)
logger.info(f"βœ… Created {len(relationships)} IN_COMMUNITY relationships")
return relationships
# STEP 9: DRIFT Search Metadata Generation
def _generate_drift_metadata(self, community_summaries: Dict[int, str], key_entities: Dict[int, List[str]]) -> Dict:
"""Generate DRIFT search metadata using existing embedding infrastructure"""
logger.info("πŸ” Generating DRIFT search metadata...")
if not self.embedding_model:
logger.warning("⚠️ Embedding model not available, skipping DRIFT metadata")
return {}
# Calculate dynamic values from actual graph data
total_communities = len(community_summaries)
total_nodes = self.nx_graph.number_of_nodes()
total_edges = self.nx_graph.number_of_edges()
avg_community_size = sum(self.community_stats.get(str(i), {}).get("member_count", 0)
for i in community_summaries.keys()) / total_communities if total_communities > 0 else 0
graph_density = total_edges / (total_nodes * (total_nodes - 1) / 2) if total_nodes > 1 else 0
# Calculate dynamic thresholds based on graph complexity
complexity_factor = min(1.0, (total_nodes + total_edges) / 10000) # Scale 0-1 based on graph size
base_confidence = 0.6 + (complexity_factor * 0.3) # Range: 0.6-0.9
base_response_time = 1.0 + (complexity_factor * 3.0) # Range: 1-4 seconds
base_memory = int(20 + (avg_community_size * complexity_factor * 5)) # Scale with size
# Adaptive configuration based on graph characteristics
max_communities_for_primer = min(total_communities, max(2, total_communities // 4))
lightweight_communities = max(1, max_communities_for_primer // 2)
standard_communities = max(2, int(max_communities_for_primer // 1.5))
comprehensive_communities = max_communities_for_primer
# Calculate dynamic iteration counts based on community distribution
max_iter = max(2, min(5, int(total_communities / 10) + 2))
hyde_count = max(2, min(5, int(avg_community_size / 5) + 2))
drift_metadata = {
"version": "1.0",
"generated_timestamp": datetime.now().isoformat(),
"configuration": {
"max_iterations": max_iter,
"confidence_threshold": round(base_confidence + 0.1, 2),
"top_k_communities": max_communities_for_primer,
"hyde_expansion_count": hyde_count,
"termination_criteria": "confidence_or_max_iterations"
},
"query_routing_config": {
"lightweight_drift": {
"triggers": ["single_entity", "simple_fact", "definition_query"],
"config": {
"primer_communities": int(lightweight_communities),
"follow_up_iterations": max(1, max_iter - 2),
"confidence_threshold": round(base_confidence, 2)
}
},
"standard_drift": {
"triggers": ["multi_entity", "relationship_query", "how_does"],
"config": {
"primer_communities": int(standard_communities),
"follow_up_iterations": max(1, max_iter - 1),
"confidence_threshold": round(base_confidence + 0.1, 2)
}
},
"comprehensive_drift": {
"triggers": ["analyze", "compare", "implications", "strategy"],
"config": {
"primer_communities": int(comprehensive_communities),
"follow_up_iterations": max_iter,
"confidence_threshold": round(base_confidence + 0.2, 2)
}
}
},
"performance_monitoring": {
"response_time_targets": {
"p50": round(base_response_time * 1.0, 1),
"p95": round(base_response_time * 2.5, 1),
"p99": round(base_response_time * 5.0, 1)
},
"resource_tracking": {
"memory_per_query": base_memory,
"cache_hit_rate_target": round(0.5 + (complexity_factor * 0.3), 2)
},
"bottleneck_identification": ["community_ranking", "follow_up_generation", "embedding_computation"]
},
"community_search_index": {},
"search_optimization": {
"total_communities": total_communities,
"avg_community_size": round(avg_community_size, 1),
"graph_density": round(graph_density, 6),
"total_nodes": total_nodes,
"total_edges": total_edges,
"max_primer_communities": max_communities_for_primer
}
}
# Process each community
for comm_id, summary in community_summaries.items():
comm_key = f"comm-{comm_id}"
try:
# Generate embeddings using existing HuggingFace model
summary_embedding = self.embedding_model.get_text_embedding(summary)
hyde_embeddings = self._generate_hyde_embeddings(summary)
follow_up_questions = self._generate_follow_up_questions(summary, comm_id, key_entities.get(comm_id, []))
# Add to search index
drift_metadata["community_search_index"][comm_key] = {
"summary": summary,
"key_entities": key_entities.get(comm_id, []),
"embeddings": {
"summary_embedding": summary_embedding,
"hyde_embeddings": hyde_embeddings
},
"follow_up_templates": follow_up_questions,
"statistics": self.community_stats.get(str(comm_id), {})
}
except Exception as e:
logger.warning(f"⚠️ Failed to generate metadata for {comm_key}: {e}")
continue
logger.info(f"βœ… Generated DRIFT metadata for {len(drift_metadata['community_search_index'])} communities")
return drift_metadata
def _generate_hyde_embeddings(self, community_summary: str) -> List[List[float]]:
"""Generate HyDE embeddings for enhanced recall"""
# Create 3 hypothetical document variations
hyde_templates = [
f"Research analysis and findings: {community_summary}",
f"Technical report and documentation: {community_summary}",
f"Business implications and strategic analysis: {community_summary}"
]
hyde_embeddings = []
for template in hyde_templates:
try:
embedding = self.embedding_model.get_text_embedding(template)
hyde_embeddings.append(embedding)
except Exception as e:
logger.warning(f"⚠️ HyDE embedding generation failed: {e}")
continue
return hyde_embeddings
def _generate_follow_up_questions(self, community_summary: str, comm_id: int, key_entities: List[str]) -> List[Dict]:
"""Generate follow-up questions using existing LLM infrastructure"""
# Professional system prompt matching Phase 1 style
system_prompt = (
"You are a specialized DRIFT search question generation assistant. Your task is to analyze community "
"summaries and generate targeted follow-up questions for iterative knowledge graph exploration.\n\n"
"CONSTITUTIONAL AI PRINCIPLES:\n"
"1. Context-Adaptive: Generate questions based on actual community content and entities\n"
"2. Search-Aware: Choose appropriate search types to guide query routing optimization\n"
"3. Relevance-First: Prioritize questions that expand understanding of community themes\n"
"4. Structured Output: Ensure consistent JSON format for programmatic consumption\n\n"
"QUESTION GENERATION GUIDELINES:\n"
"- Analyze community summary and key entities to identify knowledge gaps\n"
"- Generate questions that would reveal additional relevant information\n"
"- Use local search for entity-specific queries, relationship for connections, global for themes\n"
"- Assign relevance scores based on potential value for understanding the community\n"
"- Target entities should guide search focus and retrieval optimization"
)
user_prompt = (
f"Analyze the following community data and generate targeted follow-up questions.\n\n"
f"COMMUNITY SUMMARY:\n{community_summary}\n\n"
f"KEY ENTITIES: {', '.join(key_entities[:5]) if key_entities else 'No specific entities identified'}\n\n"
f"TASK: Generate exactly 3 strategic follow-up questions for DRIFT search.\n\n"
f"OUTPUT FORMAT (strict JSON):\n"
f"[\n"
f" {{\n"
f" \"question\": \"Specific, actionable question about the community\",\n"
f" \"relevance_score\": 0.85,\n"
f" \"search_type\": \"local\",\n"
f" \"target_entities\": [\"entity1\", \"entity2\"]\n"
f" }}\n"
f"]\n\n"
f"VALIDATION REQUIREMENTS:\n"
f"- question: Must be a clear, specific question that expands community understanding\n"
f"- relevance_score: Float 0.0-1.0 based on potential value for knowledge expansion\n"
f"- search_type: Must be one of 'local', 'relationship', or 'global'\n"
f"- target_entities: Array of relevant entity names from the key entities list\n\n"
f"IMPORTANT: Respond with ONLY the JSON array. No markdown formatting, no explanations, no code blocks."
)
try:
# Use existing LLM infrastructure
if self.llm_provider == "cerebras":
response = self._cerebras_inference(system_prompt, user_prompt)
else:
response = self._gemini_inference(system_prompt, user_prompt)
# Parse LLM response to structured questions
questions = self._parse_questions_response(response, key_entities)
return questions
except Exception as e:
logger.error(f"❌ Question generation failed for comm-{comm_id}: {e}")
return []
def _parse_questions_response(self, response: str, key_entities: List[str]) -> List[Dict]:
"""Parse LLM response into structured questions using robust multi-strategy approach"""
try:
# Calculate dynamic default relevance based on community statistics
total_nodes = self.nx_graph.number_of_nodes() if hasattr(self, 'nx_graph') else 100
node_density = min(1.0, total_nodes / 500) # Scale 0-1
default_relevance = round(0.5 + (node_density * 0.4), 2) # Range: 0.5-0.9
max_questions = max(2, min(5, len(key_entities) + 1)) # Adaptive question count
# Strategy 1: JSON array extraction with regex
try:
import re
match = re.search(r"(\[\s*\{[\s\S]*?\}\s*\])", response)
if match:
json_str = match.group(1)
try:
questions = self._smart_json_parse_questions(json_str)
if questions:
return self._validate_and_normalize_questions(questions, key_entities, default_relevance, max_questions)
except ValueError:
pass # Continue to next strategy if JSON parsing fails
except Exception:
pass
# Strategy 2: Multiple JSON objects extraction
try:
import re
pattern = r'\{[^{}]*"question"[^{}]*\}'
matches = re.findall(pattern, response)
if matches:
json_array = "[" + ",".join(matches) + "]"
try:
questions = self._smart_json_parse_questions(json_array)
if questions:
return self._validate_and_normalize_questions(questions, key_entities, default_relevance, max_questions)
except ValueError:
pass # Continue to next strategy if JSON parsing fails
except Exception:
pass
# Strategy 3: Markdown list extraction
try:
questions = self._parse_markdown_questions(response, key_entities, default_relevance)
if questions:
return self._validate_and_normalize_questions(questions, key_entities, default_relevance, max_questions)
except Exception:
pass
# Strategy 4: Generate default questions based on entities
return self._generate_default_questions(key_entities, default_relevance, max_questions)
except Exception as e:
logger.warning(f"⚠️ All question parsing strategies failed: {e}")
return self._generate_default_questions(key_entities, 0.7, 3)
def _smart_json_parse_questions(self, json_text: str) -> List[Dict]:
"""
Simple 5-step JSON parsing approach (exactly same as Phase 1)
"""
cleaned_text = json_text.strip()
# Step 1: orjson
try:
result = orjson.loads(cleaned_text.encode('utf-8'))
logger.debug("βœ… Step 1: orjson succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 1: orjson failed - {e}")
# Step 2: json-repair
try:
repaired = repair_json(cleaned_text)
result = orjson.loads(repaired.encode('utf-8'))
logger.debug("βœ… Step 2: json-repair + orjson succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 2: json-repair failed - {e}")
# Step 3: standard json
try:
result = json.loads(cleaned_text)
logger.debug("βœ… Step 3: standard json succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 3: standard json failed - {e}")
# Step 4: json-repair + standard json
try:
repaired = repair_json(cleaned_text)
result = json.loads(repaired)
logger.debug("βœ… Step 4: json-repair + standard json succeeded")
return result
except Exception as e:
logger.debug(f"❌ Step 4: json-repair + standard json failed - {e}")
# Step 5: All failed - this will trigger save failed txt files
raise ValueError("All 4 JSON parsing steps failed")
def _parse_markdown_questions(self, response: str, key_entities: List[str], default_relevance: float) -> List[Dict]:
"""Parse questions from markdown or plain text format"""
questions = []
# Look for numbered lists or bullet points
import re
patterns = [
r'\d+\.\s*(.+?)(?=\n\d+\.|\n-|\n\*|$)', # Numbered list
r'-\s*(.+?)(?=\n-|\n\*|\n\d+\.|$)', # Dash list
r'\*\s*(.+?)(?=\n\*|\n-|\n\d+\.|$)' # Asterisk list
]
for pattern in patterns:
matches = re.findall(pattern, response, re.MULTILINE | re.DOTALL)
if matches and len(matches) >= 2:
for i, match in enumerate(matches[:5]): # Max 5 questions
question_text = match.strip().replace('\n', ' ')
if len(question_text) > 10: # Reasonable question length
search_type = 'global' if any(word in question_text.lower()
for word in ['analyze', 'compare', 'overall', 'trends']) else 'local'
questions.append({
'question': question_text,
'relevance_score': max(0.6, default_relevance - (i * 0.1)),
'search_type': search_type,
'target_entities': key_entities[:2] if key_entities else []
})
break
return questions
def _generate_default_questions(self, key_entities: List[str], default_relevance: float, max_questions: int) -> List[Dict]:
"""Generate default questions when parsing fails"""
if not key_entities:
return []
# Template questions based on entity analysis
question_templates = [
("What is {entity} and what role does it play?", "local"),
("How does {entity} relate to other entities in this community?", "relationship"),
("What are the key characteristics and properties of {entity}?", "local"),
("What trends or patterns involve {entity}?", "global"),
("How might {entity} impact the broader context?", "global")
]
questions = []
entities_to_use = key_entities[:max_questions]
for i, entity in enumerate(entities_to_use):
if i < len(question_templates):
template, search_type = question_templates[i]
question = template.format(entity=entity)
questions.append({
'question': question,
'relevance_score': max(0.6, default_relevance - (i * 0.05)),
'search_type': search_type,
'target_entities': [entity]
})
return questions
def _validate_and_normalize_questions(self, questions: List[Dict], key_entities: List[str],
default_relevance: float, max_questions: int) -> List[Dict]:
"""Validate and normalize question format"""
normalized = []
for q in questions:
if not isinstance(q, dict):
continue
# Extract question text
question = q.get('question') or q.get('q') or q.get('text')
if not question or len(str(question).strip()) < 5:
continue
# Extract and validate relevance score
relevance = q.get('relevance_score', default_relevance)
try:
relevance = float(relevance)
if relevance <= 0 or relevance > 1:
relevance = default_relevance
except (ValueError, TypeError):
relevance = default_relevance
# Extract and validate search type
search_type = q.get('search_type', 'local')
if search_type not in ('local', 'relationship', 'global'):
search_type = 'local'
# Extract target entities
target_entities = q.get('target_entities', [])
if not isinstance(target_entities, list):
target_entities = []
# Ensure we have some target entities
if not target_entities and key_entities:
target_entities = key_entities[:2]
normalized.append({
'question': str(question).strip(),
'relevance_score': round(relevance, 2),
'search_type': search_type,
'target_entities': target_entities
})
if len(normalized) >= max_questions:
break
return normalized
# STEP 10: Main Processing Entry Point
def generate_summaries(self, input_path: str = None, output_path: str = None) -> bool:
"""Main entry point for Phase 3"""
if output_path is None:
output_path = "workspace/graph_data/graph-data-final.json"
logger.info("πŸš€ Starting Phase 3: Community Summarization")
logger.info("=" * 60)
start_time = time.time()
# Step 1: Load Phase 2 output
if not self.load_graph_data(input_path):
return False
# Step 2: Build NetworkX graph
self.nx_graph = self._build_networkx_graph()
# Step 3: Extract community assignments
self.community_assignments = self._extract_community_assignments()
# Step 4: Generate LLM summaries
community_summaries = self._generate_community_summaries()
# Step 5: Identify key entities
key_entities = self._identify_key_entities()
# Step 6: Create community nodes
community_nodes = self._create_community_nodes(community_summaries, key_entities)
# Step 7: Create IN_COMMUNITY relationships
community_relationships = self._create_in_community_relationships(community_nodes)
# Step 8: Merge everything
self.graph_data["nodes"].extend(community_nodes)
self.graph_data["relationships"].extend(community_relationships)
# Step 9: Add communities section
self.graph_data["communities"] = {
"algorithm": "Leiden",
"total_communities": len(community_summaries),
"modularity_score": self.graph_data["metadata"]["community_detection"]["modularity_score"],
"summaries": {
f"comm-{k}": v for k, v in community_summaries.items()
}
}
# Step 10: Generate DRIFT search metadata
drift_metadata = self._generate_drift_metadata(community_summaries, key_entities)
if drift_metadata:
self.graph_data["drift_search_metadata"] = drift_metadata
logger.info("βœ… Added DRIFT search metadata to graph data")
# Step 11: Clean up temporary data
if "community_stats" in self.graph_data:
del self.graph_data["community_stats"]
# Step 12: Update metadata
self.graph_data["metadata"]["phase"] = "final"
self.graph_data["metadata"]["entity_count"] = len([n for n in self.graph_data["nodes"] if "Community" not in n["labels"]])
self.graph_data["metadata"]["community_count"] = len(community_nodes)
self.graph_data["metadata"]["total_node_count"] = len(self.graph_data["nodes"])
self.graph_data["metadata"]["total_relationship_count"] = len(self.graph_data["relationships"])
# Step 13: Save final output
if self._save_final_output(output_path):
elapsed = time.time() - start_time
logger.info("=" * 60)
logger.info(f"βœ… Phase 3 completed successfully in {elapsed:.1f}s")
logger.info("πŸ“Š Final stats:")
logger.info(f" - Total nodes: {len(self.graph_data['nodes'])}")
logger.info(f" - Entity nodes: {self.graph_data['metadata']['entity_count']}")
logger.info(f" - Community nodes: {len(community_nodes)}")
logger.info(f" - Total relationships: {len(self.graph_data['relationships'])}")
logger.info(f" - Communities with summaries: {len(community_summaries)}")
logger.info(f" - Output saved to: {output_path}")
return True
else:
return False
# STEP 14: Save Final Output
def _save_final_output(self, output_path: str) -> bool:
"""Save graph-data-final.json with DRIFT search metadata"""
try:
# Ensure output directory exists
output_dir = Path(output_path).parent
output_dir.mkdir(parents=True, exist_ok=True)
# Save final output
with open(output_path, 'w', encoding='utf-8') as f:
json.dump(self.graph_data, f, indent=2, ensure_ascii=False)
# Calculate file size
output_size = os.path.getsize(output_path)
output_size_mb = output_size / (1024 * 1024)
logger.info(f"πŸ’Ύ Saved final output: {output_path} ({output_size_mb:.2f} MB)")
return True
except Exception as e:
logger.error(f"❌ Error saving final output: {e}")
return False
# STEP 15: Main Entry Point
def main():
"""Main function to run Phase 3: Community Summarization with DRIFT Search Metadata"""
logger.info("πŸš€ GraphRAG Phase 3: Community Summarization + DRIFT Search Metadata")
logger.info(" Input: graph-data-phase-2.json (from Phase 2)")
logger.info(" Output: graph-data-final.json (with DRIFT search metadata)")
logger.info("")
# Choose LLM provider from environment or default to cerebras
llm_provider = os.getenv("GRAPH_LLM_PROVIDER", "cerebras").lower()
logger.info(f" Using LLM provider: {llm_provider.upper()}")
try:
# Initialize Phase 3 processor
processor = GraphBuilderPhase3(llm_provider=llm_provider)
# Generate summaries
success = processor.generate_summaries()
if success:
logger.info("")
logger.info("βœ… Phase 3 completed successfully!")
logger.info("οΏ½ DRIFT search metadata generated and included")
logger.info("οΏ½πŸ“‹ Next step: Upload to Neo4j using 3b_save_to_graph_db.py")
logger.info(" The graph-data-final.json is now ready for Neo4j import with DRIFT capabilities")
return 0
else:
logger.error("")
logger.error("❌ Phase 3 failed")
logger.error(" Please check the logs above for details")
return 1
except Exception as e:
logger.error(f"❌ Phase 3 pipeline failed: {e}")
import traceback
logger.error(traceback.format_exc())
return 1
if __name__ == "__main__":
exit(main())