bosprimigenious opened a new issue, #397:
URL: https://github.com/apache/hugegraph-ai/issues/397
### Bug Type
data inconsistency in the local PropertyGraphExtract operator
### Before submit
- [x] Searched open/closed issues, PRs and relevant comments for
numeric/zero/falsy/custom IDs. #331/#361 concern earlier normalization/import
property changes; their current parser still reproduces this. No overlapping
parser fix found. This is independent of the Ontogeny proposal #376 and
implementations #380/#391.
### Environment
- hugegraph-ai main: 3cea006825b49f7a66e47d9926501b1060069e7b
- Python 3.12.6, macOS arm64; uv-managed local operator/test dependencies
from the declared manifest ranges; no tracked uv.lock on this base
- No graph server or real model call. The deterministic text-producing
object below only supplies JSON to the actual operator.
### Expected & Actual behavior
Given vertices 0 and 1 and an edge 0 -> 1, PropertyGraphExtract.run keeps
both vertices but silently drops the edge. Integer 0 is tested with truthiness
in the vertex map, endpoint resolution and normalized-edge filter. The same
issue prevents a model-generated numeric alias 0 from resolving to a
PRIMARY_KEY canonical ID.
CUSTOMIZE_NUMBER validation in HugeGraph GraphTransaction.checkId requires a
non-null numeric ID and LongId stores the long value; this is source-contract
evidence, not a claim of live server insertion.
### Vertex/Edge example and Schema
Run from the module development environment:
```python
import json
import socket
from types import SimpleNamespace
connections=[]
def reject_connect(self,address):
connections.append(repr(address))
raise AssertionError("No network allowed in the reproduction")
socket.socket.connect=reject_connect
from hugegraph_llm.operators.llm_op.property_graph_extract import
PropertyGraphExtract
schema={"vertexlabels":[{"name":"person","id_strategy":"CUSTOMIZE_NUMBER","primary_keys":[],"properties":[],"nullable_keys":[]}],"edgelabels":[{"name":"knows","source_label":"person","target_label":"person","properties":[]}]}
graph={"vertices":[{"id":0,"label":"person","properties":{}},{"id":1,"label":"person","properties":{}}],"edges":[{"label":"knows","outV":0,"outVLabel":"person","inV":1,"inVLabel":"person","properties":{}}]}
response=json.dumps(graph)
model=SimpleNamespace(generate=lambda **kwargs: response)
result=PropertyGraphExtract(llm=model).run({"schema":schema,"chunks":["Vertex
zero knows vertex one."]})
print(json.dumps({"vertices":result["vertices"],"edges":result["edges"],"call_count":result["call_count"],"network_attempts":connections},indent=2))
assert len(result["edges"])==1, "A valid edge from CUSTOMIZE_NUMBER vertex 0
was dropped"
assert result["edges"][0]["outV"]==0 and type(result["edges"][0]["outV"]) is
int
```
Actual: `vertices` retains ID 0 and ID 1, `edges` is `[]`, then the
assertion fails. Expected: one edge whose outV is integer 0 and inV is integer
1. The socket guard records zero connections.
### Proposed narrow fix and verification
Treat integer 0 as present in the existing ID-normalization path while
preserving rejection of None, empty strings and False. No model/API/dependency
or SDK changes. Ten regressions give 4 failures/6 passes on the original
source; after the fix, the complete operator test file gives 46 passes (one
existing Pydantic warning). Seven extra local parser boundary controls also
pass. Whole-repository Ruff format/check and scoped ty pass.
Full acceptance is NOT READY: an additional HTTP integration collection
fails because the lightweight environment lacks fastapi; the full application
environment, coverage/matrix, live HugeGraph insertion and provider tests are
unverified. In particular this does not fix the separate existing numeric-ID
write-through limitation in CommitToHugeGraph.
AI disclosure: reproduction, patch and this text were prepared with OpenAI
Codex. No human review or end-to-end graph/provider verification is claimed.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]