File size: 9,150 Bytes
fa9ea37 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 | """
Graph RAG — Knowledge Graph Builder
Builds a NetworkX graph where:
- Nodes = chunks (from doc1 & doc2)
- Edges = relationships between chunks:
* sequential : consecutive chunks in same document
* same_section : chunks sharing the same heading/section
* cross_similar: high cosine similarity between doc1 chunk & doc2 chunk
* entity_link : chunks sharing important noun phrases (entities)
"""
import re
import networkx as nx
from typing import List, Dict, Any, Tuple
from sentence_transformers import SentenceTransformer
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
from .chunker import Chunk
_EMBED_MODEL_NAME = "all-MiniLM-L6-v2"
_CROSS_SIM_THRESHOLD = 0.55 # min similarity to create a cross-doc edge
_ENTITY_MIN_LEN = 4 # min characters for an entity term
def _extract_noun_phrases(text: str) -> set:
"""
Lightweight noun phrase extraction via regex patterns.
No spacy dependency — works in constrained environments.
"""
# Capitalised multi-word phrases and key technical terms
patterns = [
r'\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+)+\b', # "Neural Network", "New York"
r'\b[A-Z]{2,}\b', # acronyms: "RAG", "LLM"
r'\b\w{5,}\b', # any long word (catch technical terms)
]
entities = set()
for pat in patterns:
found = re.findall(pat, text)
entities.update(f.strip().lower() for f in found if len(f) >= _ENTITY_MIN_LEN)
# Remove very common stopwords
stopwords = {'which', 'these', 'those', 'their', 'there', 'where', 'about',
'would', 'could', 'should', 'other', 'being', 'using', 'having'}
return entities - stopwords
class GraphBuilder:
"""
Builds and queries a knowledge graph from doc chunks.
"""
def __init__(self):
self._model = SentenceTransformer(_EMBED_MODEL_NAME)
self.graph: nx.Graph = nx.Graph()
self._chunk_map: Dict[str, Chunk] = {} # chunk_id -> Chunk
# ------------------------------------------------------------------
# Build
# ------------------------------------------------------------------
def build(self, doc1_chunks: List[Chunk], doc2_chunks: List[Chunk]) -> nx.Graph:
"""
Full graph construction pipeline.
Returns the built NetworkX graph.
"""
self.graph = nx.Graph()
self._chunk_map = {}
all_chunks = doc1_chunks + doc2_chunks
# 1. Add nodes
for chunk in all_chunks:
self._chunk_map[chunk.chunk_id] = chunk
self.graph.add_node(
chunk.chunk_id,
text=chunk.text[:200], # store snippet
doc_id=chunk.doc_id,
section=chunk.section,
chunk_index=chunk.chunk_index,
entities=list(_extract_noun_phrases(chunk.text)),
)
# 2. Sequential edges (within same doc)
self._add_sequential_edges(doc1_chunks)
self._add_sequential_edges(doc2_chunks)
# 3. Same-section edges
self._add_section_edges(all_chunks)
# 4. Cross-document similarity edges
self._add_cross_similarity_edges(doc1_chunks, doc2_chunks)
# 5. Entity co-occurrence edges
self._add_entity_edges(all_chunks)
return self.graph
def _add_sequential_edges(self, chunks: List[Chunk]) -> None:
sorted_chunks = sorted(chunks, key=lambda c: c.chunk_index)
for i in range(len(sorted_chunks) - 1):
a, b = sorted_chunks[i], sorted_chunks[i + 1]
self.graph.add_edge(
a.chunk_id, b.chunk_id,
relation="sequential",
weight=0.9,
)
def _add_section_edges(self, chunks: List[Chunk]) -> None:
section_map: Dict[str, List[str]] = {}
for chunk in chunks:
key = f"{chunk.doc_id}::{chunk.section}"
section_map.setdefault(key, []).append(chunk.chunk_id)
for ids in section_map.values():
for i in range(len(ids)):
for j in range(i + 1, len(ids)):
if not self.graph.has_edge(ids[i], ids[j]):
self.graph.add_edge(
ids[i], ids[j],
relation="same_section",
weight=0.6,
)
def _add_cross_similarity_edges(
self, doc1_chunks: List[Chunk], doc2_chunks: List[Chunk]
) -> None:
if not doc1_chunks or not doc2_chunks:
return
texts1 = [c.text for c in doc1_chunks]
texts2 = [c.text for c in doc2_chunks]
emb1 = self._model.encode(texts1, batch_size=32, show_progress_bar=False)
emb2 = self._model.encode(texts2, batch_size=32, show_progress_bar=False)
sim_matrix = cosine_similarity(emb1, emb2)
for i, c1 in enumerate(doc1_chunks):
for j, c2 in enumerate(doc2_chunks):
sim = float(sim_matrix[i, j])
if sim >= _CROSS_SIM_THRESHOLD:
self.graph.add_edge(
c1.chunk_id, c2.chunk_id,
relation="cross_similar",
weight=round(sim, 4),
similarity=round(sim, 4),
)
def _add_entity_edges(self, chunks: List[Chunk]) -> None:
entity_to_chunks: Dict[str, List[str]] = {}
for chunk in chunks:
entities = _extract_noun_phrases(chunk.text)
for ent in entities:
entity_to_chunks.setdefault(ent, []).append(chunk.chunk_id)
for ent, ids in entity_to_chunks.items():
if len(ids) < 2:
continue
# Only connect cross-doc pairs to avoid too many same-doc entity edges
doc_ids = {self._chunk_map[cid].doc_id: cid for cid in ids}
if len(doc_ids) >= 2:
cids = list(doc_ids.values())
for i in range(len(cids)):
for j in range(i + 1, len(cids)):
if not self.graph.has_edge(cids[i], cids[j]):
self.graph.add_edge(
cids[i], cids[j],
relation="entity_link",
entity=ent,
weight=0.5,
)
# ------------------------------------------------------------------
# Query
# ------------------------------------------------------------------
def retrieve(
self,
query: str,
seed_chunks: List[Dict[str, Any]], # from VectorStore.search()
hops: int = 2,
max_nodes: int = 10,
) -> List[Dict[str, Any]]:
"""
Graph-aware retrieval:
1. Start from seed chunk nodes (vector search results)
2. Expand via BFS up to `hops` hops, prioritising high-weight edges
3. Return unique chunks from both docs, ranked by relevance
"""
visited = set()
result_nodes = []
seed_ids = [
f"{s['doc_id']}_chunk_{s['chunk_index']}"
for s in seed_chunks
if s.get('chunk_index') is not None
]
# BFS queue: (node_id, remaining_hops, accumulated_weight)
queue = [(nid, hops, 1.0) for nid in seed_ids if nid in self.graph]
while queue and len(result_nodes) < max_nodes:
node_id, remaining, acc_weight = queue.pop(0)
if node_id in visited:
continue
visited.add(node_id)
chunk = self._chunk_map.get(node_id)
if chunk:
result_nodes.append({
"chunk_id": node_id,
"text": chunk.text,
"doc_id": chunk.doc_id,
"section": chunk.section,
"relevance": round(acc_weight, 4),
})
if remaining > 0:
neighbors = sorted(
self.graph[node_id].items(),
key=lambda x: x[1].get("weight", 0),
reverse=True,
)
for neighbor_id, edge_data in neighbors[:4]: # top-4 neighbours
if neighbor_id not in visited:
queue.append((
neighbor_id,
remaining - 1,
acc_weight * edge_data.get("weight", 0.5),
))
# Sort by relevance
result_nodes.sort(key=lambda x: x["relevance"], reverse=True)
return result_nodes[:max_nodes]
def get_stats(self) -> Dict[str, Any]:
edge_types = {}
for _, _, data in self.graph.edges(data=True):
rel = data.get("relation", "unknown")
edge_types[rel] = edge_types.get(rel, 0) + 1
return {
"nodes": self.graph.number_of_nodes(),
"edges": self.graph.number_of_edges(),
"edge_types": edge_types,
} |