semantica

Set ALLOWED_ORIGINS (or EXPLORER_CORS_ORIGINS) to include the exact browser origin before starting the Explorer. By default, only http://localhost:5173 and http://127.0.0.1:5173 are allowed, and the same allowlist gates the /ws/graph-updates WebSocket handshake.

For port 8020, run:

ALLOWED_ORIGINS="http://127.0.0.1:8020,http://localhost:8020" \
semantica-explorer --graph any_graph.json --port 8020

Then open http://127.0.0.1:8020 and restart the server after changing the variable. This issue was observed in Semantica 0.6.7.

Install the optional dependency, or skip the Node2Vec tests in slim-core environments.

For a full test environment:

pip install -e ".[dev,graph-embeddings]"

This installs gensim>=4.4.0. With it installed, all affected Node2Vec tests should pass.

If gensim is intentionally absent, add this guard at the top of each affected test module:

pytest.importorskip("gensim")  # Node2Vec requires semantica[graph-embeddings]

Apply it to the eight tests in:
- tests/context/test_banking_context_graphs_e2e.py
- tests/context/test_context_explainability_regression.py
- tests/kg/test_enhanced_algorithms_e2e.py
- tests/kg/test_provenance_integration.py
- tests/test_030_realworld_comprehensive.py
- tests/vector_store/test_decision_embedding_pipeline.py

Do not move the gensim check from NodeEmbedder.__init__ into _train_word2vec(): the slim-core contract intentionally relies on the constructor raising so callers can disable Node2Vec gracefully. The tests requiring this behavior are test_node_embedder_gensim_missing_hint, test_agent_context_with_graph_builds_without_gensim, and test_graph_components_without_gensim_keep_everything_but_node2vec.

Install the optional dependency in your virtualenv:

pip install -e ".[graph-embeddings]"

or, for a released package:

pip install "semantica[graph-embeddings]"

Then rerun:

python -m pytest tests/kg/test_node_embeddings.py -q

For the repository test fix, do not add a module-level pytest.importorskip("gensim") because test_init_gensim_unavailable and test_init_invalid_method must run without gensim. Instead:

1. Register a marker in pyproject.toml:

[tool.pytest.ini_options]
markers = [
    "no_gensim: test runs without the optional gensim dependency"
]

2. Add an autouse fixture that skips gensim-dependent tests unless the test is marked no_gensim:

import pytest
from semantica.kg.node_embeddings import GENSIM_AVAILABLE

@pytest.fixture(autouse=True)
def require_gensim(request):
    if not GENSIM_AVAILABLE and "no_gensim" not in request.keywords:
        pytest.skip("Install semantica[graph-embeddings] to run Node2Vec tests")

3. Mark test_init_gensim_unavailable and test_init_invalid_method with @pytest.mark.no_gensim.

Expected result without gensim: 2 passed, 31 skipped. With gensim installed: all 33 tests pass.

CSVExporter.export_relationships() only recognizes source/source_id and target/target_id as endpoint field names. Backend-shaped records use start_node_id/end_node_id (Neo4j, FalkorDB, Neptune, and renamed AGE edges) or start_id/end_id (raw AGE), so the normalizing path silently writes empty endpoint columns with no error or warning. Fixes/workarounds:

1. **Fix (patch)**: In semantica/export/csv_exporter.py (export_relationships), run records through canonical_endpoints from semantica/export/endpoint_names.py (which defines SOURCE_ENDPOINT_NAMES/TARGET_ENDPOINT_NAMES) at the start of the method — the same approach the Parquet and Arrow exporters already use. This maps all backend spellings into source_id/target_id while keeping source/target and source_id/target_id working as before.

2. **Workaround (no code change)**: Rename the fields before exporting:

record = {"id": "r1", "start_node_id": "alice", "end_node_id": "acme", "type": "WORKS_AT"}
record["source_id"], record["target_id"] = record.pop("start_node_id"), record.pop("end_node_id")
CSVExporter().export_relationships([record], "rel.csv")

3. **Alternative path**: export({"edges": [record]}) preserves the values but writes them under the original column names (start_node_id/end_node_id) without normalization. Parquet and Arrow exporters are not affected.

FileObject instances are silently skipped because GraphBuilder._process_item() does not recognize them; it expects text strings or dictionaries such as {"text": ...}. As a workaround, convert each ingested FileObject before building the graph:

docs = FileIngestor().ingest_directory("./docs/", recursive=True)
kg = GraphBuilder(merge_entities=True, enable_temporal=True).build(
    [{"text": d.text} for d in docs]
)

The permanent fix in PR #1951 makes build() extract text from FileObject.text directly and logs a warning naming unsupported source types instead of skipping them silently. Update to a release containing PR #1951, or apply that change if running from main.