Light GraphRAG (#4585)

### What problem does this PR solve?

#4543

### Type of change

- [x] New Feature (non-breaking change which adds functionality)
This commit is contained in:
Kevin Hu
2025-01-22 19:43:14 +08:00
committed by GitHub
parent 1a367664f1
commit dd0ebbea35
55 changed files with 5461 additions and 4000 deletions

View File

@ -59,7 +59,7 @@ class Dealer:
if key in req and req[key] is not None:
condition[field] = req[key]
# TODO(yzc): `available_int` is nullable however infinity doesn't support nullable columns.
for key in ["knowledge_graph_kwd", "available_int"]:
for key in ["knowledge_graph_kwd", "available_int", "entity_kwd", "from_entity_kwd", "to_entity_kwd", "removed_kwd"]:
if key in req and req[key] is not None:
condition[key] = req[key]
return condition
@ -198,6 +198,11 @@ class Dealer:
return answer, set([])
ans_v, _ = embd_mdl.encode(pieces_)
for i in range(len(chunk_v)):
if len(ans_v[0]) != len(chunk_v[i]):
chunk_v[i] = [0.0]*len(ans_v[0])
logging.warning("The dimension of query and chunk do not match: {} vs. {}".format(len(ans_v[0]), len(chunk_v[i])))
assert len(ans_v[0]) == len(chunk_v[0]), "The dimension of query and chunk do not match: {} vs. {}".format(
len(ans_v[0]), len(chunk_v[0]))
@ -434,6 +439,8 @@ class Dealer:
es_res = self.dataStore.search(fields, [], condition, [], OrderByExpr(), p, bs, index_name(tenant_id),
kb_ids)
dict_chunks = self.dataStore.getFields(es_res, fields)
for id, doc in dict_chunks.items():
doc["id"] = id
if dict_chunks:
res.extend(dict_chunks.values())
if len(dict_chunks.values()) < bs: