test_retry_logic.py injected sys.modules["openai"] = MagicMock() at module
level so providers.py could be imported without the real openai package.
Those mocks were never restored, leaving openai (and spacy, instructor etc.)
as MagicMock objects for the entire test session. This caused
test_pr482_deepseek_openai tests to receive a MagicMock when importing
openai.OpenAI, making MagicMock(spec=OpenAI) raise InvalidSpecError.
Fix: save original sys.modules entries before injection and restore them
immediately after the semantica imports that needed the mocks complete.
The mock objects remain bound inside the already-imported provider module,
so test_retry_logic tests are unaffected; other test modules now see the
real packages again.
Co-authored-by: Zohaib Hassan <zohaib179949@gmail.com>
Co-authored-by: KaifAhmad1 <kaifahmad087@gmail.com>
subprocess.CompletedProcess[str] as a return annotation is not subscriptable
at runtime on Python 3.8, causing test collection to abort before any tests
run. Adding PEP 563 deferred evaluation makes all annotations strings at
import time, restoring 3.8 compatibility without changing behaviour on 3.9+.
Co-authored-by: Zohaib Hassan <zohaib179949@gmail.com>
Co-authored-by: KaifAhmad1 <kaifahmad087@gmail.com>
- Replace deepseek.Client with openai.OpenAI(base_url="https://api.deepseek.com/v1")
in DeepSeekProvider._init_client(); the deepseek PyPI package has no Client class
- Add self.base_url = "https://api.deepseek.com/v1" to DeepSeekProvider.__init__()
(missing from original PR; caused AttributeError on every instantiation)
- Fix verbose_mode NameError in BaseProvider.generate_typed() instructor path
- Update pyproject.toml: llm-deepseek extra now declares openai>=1.0.0
- Update _init_client warning message to reference openai library
- Add 19 tests in tests/semantic_extract/test_pr482_deepseek_openai.py
- Update CHANGELOG.md
Co-authored-by: liling <liling@users.noreply.github.com>
Co-authored-by: KaifAhmad1 <mohammadk78600@gmail.com>
- Add `extract_temporal_bounds: bool = False` to `extract_relations_llm()`.
When True the LLM prompt is extended with a calibrated confidence scale
and four few-shot examples; each returned Relation gains valid_from,
valid_until, temporal_confidence, and temporal_source_text in metadata.
Low confidence (<0.5) with non-null dates logs a WARNING. Default False
preserves 100% backward compatibility.
- Add `RelationWithTemporalOut` / `RelationsWithTemporalResponse` Pydantic
schemas so the four temporal fields are captured from structured LLM
output (separate from RelationOut which uses extra="ignore").
- New `semantica/kg/temporal_normalizer.py` — `TemporalNormalizer` class
(zero LLM calls, pure regex + dateutil arithmetic):
* normalize(value) → (start, end) UTC datetimes or None
* Resolution order: ISO 8601 → partial dates (year/month/Q) →
ambiguity detection → domain phrase map → relative phrases
* normalize_phrase(phrase) → metadata dict or None
* Default phrase map covers 13 domains: General, Policy, Healthcare,
Drug Discovery, Cybersecurity, Supply Chain, Finance, Energy
* TemporalAmbiguityWarning for DD/MM/YYYY-style ambiguous inputs
* Custom phrase_map at construction (merged over defaults)
- Add `TemporalAmbiguityWarning(UserWarning)` to exceptions.py.
- Export `TemporalNormalizer` from `semantica/kg/__init__.py`.
- Propagate `extract_temporal_bounds` through `_extract_relations_chunked`
and add flag to cache key to prevent cross-mode cache pollution.
- 53 new tests in tests/semantic_extract/test_temporal_extraction.py;
zero real LLM calls, suite runs in ~3.5s. 873 existing tests unaffected.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Closes#408
Previously `_init_client` assigned the raw `ollama` module to
`self.client`, so the `base_url` parameter was silently ignored and
every request hit the default localhost:11434. Now an `ollama.Client`
instance is created with `host=self.base_url`, so remote Ollama servers
are reachable.
Three regression tests added to prevent recurrence:
- default base_url is forwarded as host
- custom base_url (e.g. http://192.168.1.3:11434) is forwarded as host
- self.client is never the raw module
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Bug 1 — _parse_relation_result (methods.py):
Relations whose subject/object weren't in the pre-extracted NER list were
silently dropped because match_entity() returned None and the old code
gated on `if subject_entity and object_entity`. Now unmatched names
produce a synthetic UNKNOWN Entity so every LLM-returned relation is
preserved (all three Apple co-founders are now returned).
Bug 2 — _match_pattern (reasoner.py):
Rewrote the regex builder to split on ?var placeholders first, then
apply re.escape() only to the surrounding literal segments. The old
approach (escape-then-sub) left edge cases where pre-bound variables
and multi-word values with spaces could fail to unify. The new
implementation also handles repeated variables via backreferences and
uses non-greedy .+? to avoid over-consuming literal separators.
Closes#354
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Verify that temperature parameter is omitted from API calls when None,
allowing models to use their defaults. Tests cover OpenAI, Groq, Gemini,
Ollama, and DeepSeek providers.
- Implemented high-throughput parallel batch processing across all core extractors (NERExtractor, RelationExtractor, TripletExtractor, EventDetector, SemanticNetworkExtractor) using ThreadPoolExecutor.
- Added max_workers configuration parameter (default: 1) to all extractor extract() methods.
- Implemented parallel processing for large document chunking in _extract_entities_chunked and _extract_relations_chunked.
- Enhanced ProgressTracker to be thread-safe.
- Optimized setUpClass in tests to reduce Groq LLM initialization overhead.
- Updated documentation and usage examples.
- Implemented ML/LLM -> Pattern -> Last Resort fallback chains for NER, Relation, and Triplet extractors to prevent empty results.
- Added provenance metadata (batch_index, document_id) to all extraction schemas (Entity, Relation, Triplet, etc.).
- Unified batch processing API with progress tracking across all extractors.
- Updated documentation (module usage and reference docs) to reflect new features.
- Added robustness and batch provenance tests.
- Robust ID extraction in CentralityCalculator, CommunityDetector, and ConnectivityAnalyzer
- Support for direct Entity objects and dictionaries as node identifiers
- Improved Entity hashability in utils/types.py
- Added integration test to verify fix and prevent regression
