summaryrefslogtreecommitdiffstats
path: root/searchlib/src/tests/features
Commit message (Collapse)AuthorAgeFilesLines
* Make DirectTensorAttribute::set_tensor member function private.Tor Egge2023-03-021-1/+1
|
* Test closest feature setup failures.Tor Egge2023-02-271-1/+60
|
* Rename f2 to f when appropriate.Tor Egge2023-02-271-15/+15
| | | | | Only allow mixed tensor with 1 sparse and 1 dense dimension. Return empty tensor if any distance calculator is missing.
* Add closest feature.Tor Egge2023-02-272-0/+159
|
* Use gtest for closeness and distance features unit tests.Tor Egge2023-02-272-53/+138
|
* - Improve sanity checking of input to distance-to-path feature.Henning Baldersheim2023-02-061-4/+5
| | | | - Keep comment
* Deinline large destructors and clean up some code based on clion hints.Henning Baldersheim2023-02-011-0/+3
|
* GC some compiler pragmas not needed.Henning Baldersheim2023-01-301-38/+3
|
* Expose SameElement query terms to ranking.Geir Storli2023-01-121-0/+11
| | | | | | A TermFieldMatchData is allocated per SameElement term, and this is used to signal matching docids in doUnpack() on the SameElement search iterator. This allows using the matches() rank feature on a field (virtual) that is searched using a SameElement term.
* Change from typedef to using in searchlib C++ code.Geir Storli2022-12-217-29/+29
|
* Separate where you nede an actual hash, and where you want the string hashed ↵Henning Baldersheim2022-11-201-10/+10
| | | | to a double.
* Use AttributeBuilder in prod features unit tests.Geir Storli2022-10-313-79/+51
|
* Revert "Revert "Unify access to assets needed during rank-setup.""Henning Baldersheim2022-09-071-11/+11
|
* Revert "Unify access to assets needed during rank-setup."Tor Egge2022-09-071-11/+11
|
* Unify access to assets needed during rank-setup.Henning Baldersheim2022-09-061-11/+11
|
* Test lazy distance calculation in distance and closeness features.Geir Storli2022-07-112-3/+60
|
* Share fixture classes used by distance and closeness unit tests.Geir Storli2022-07-084-186/+24
|
* more testing (arnej)Håvard Pettersen2022-06-091-3/+5
|
* disallow bad position fieldArne H Juul2022-06-091-2/+2
|
* Add bm25() and matches() and remove now() as default rank features to be ↵Geir Storli2022-06-082-2/+8
| | | | dumped on Vespa 8.
* Include searchcommon/attribut/config.h yourself and prepare for hiding large ↵Henning Baldersheim2022-05-258-8/+9
| | | | | | implementation. Add setters to vespalib::GrowStrategy.
* support default tensor values for query featureHåvard Pettersen2022-05-131-0/+15
|
* Always use MultiValueReadView for dot product feature on array attributes.Tor Egge2022-04-211-10/+42
|
* Use IMultiValueReadView in dot product feature instead of getRawValues().Tor Egge2022-04-111-2/+2
|
* use getopt/getopt_long directlyHåvard Pettersen2022-04-011-6/+5
|
* Remove inlining warnings (searchlib).Tor Egge2022-02-268-0/+27
|
* add unit tests for new featuresArne H Juul2022-01-312-1/+67
|
* auto-unbox scalar results from interpreted ranking expressionsHåvard Pettersen2022-01-271-2/+6
|
* added new rank feature "onnx(model_name)"Håvard Pettersen2021-11-021-6/+23
| | | | | | | | | | | Works the same as "onnxModel(model_name)", but is not treated as the exact same feature (features are currently not allowed to have multiple base names). If both variants are used at the same time, the model may be calculated twice, but the model cache will still make sure that the model itself is only loaded once. The plan is to deprecate and possibly remove the "onnxModel(model_name)" variant at some point in the future.
* Update 2019 Oath copyrights.gjoranv2021-10-272-2/+2
|
* Don't inline OnnxFeatureTest destructor.Tor Egge2021-10-271-0/+3
|
* report onnx model evaluation as an issueHåvard Pettersen2021-10-131-0/+14
|
* Update Verizon Media copyright notices.gjoranv2021-10-078-8/+8
|
* Update 2017 copyright notices.gjoranv2021-10-0750-50/+50
|
* avoid crash on run-time onnx errorsHåvard Pettersen2021-08-113-9/+12
| | | | | | | | - warn about onnx model dry-run being disabled - catch and report onnx errors during ranking - zero-fill failed results to avoid re-using previous results - use explicit output size in fragile model (output became float[2] instead of float[batch] anyways)
* dry run onnx models on setupHåvard Pettersen2021-06-234-3/+78
|
* also test tensorFromLabels with single-value integer attributeArne Juul2021-06-181-0/+18
|
* use explicit expression nameHåvard Pettersen2021-06-021-11/+11
|
* external ranking expressionsHåvard Pettersen2021-05-201-6/+23
| | | | loaded from potentially compressed files
* Revert "external ranking expressions"Harald Musum2021-05-201-23/+6
|
* external ranking expressionsHåvard Pettersen2021-05-181-6/+23
| | | | loaded from potentially compressed files
* support functional constantsHåvard Pettersen2021-05-111-1/+45
| | | | | | | | | | | | be more specific about const number values (GDBT trees): is_const -> is_const_double get_const_value -> get_const_double_value Add more generic 'get_const_value' that can also be used for tensor values. Allow it to be called even in the case of parse error, in which case it does not produce a value (same as for non-const sub-expressions).
* Use int for FastOS_UNIX_Application::GetOpt() return value.Tor Egge2021-04-301-1/+1
|
* all scalars must be doubleHåvard Pettersen2021-03-042-3/+3
| | | | and all operation results must be at least float
* - Add the cost to expensive tests.Henning Baldersheim2021-01-131-1/+1
| | | | - Use less resources on the fnet_parallel_rpc test bu just using as many threads as you have cores.
* use SimpleValue instead of EngineOrFactoryArne Juul2020-12-025-31/+34
|
* avoid using vespalib::tensor directlyArne Juul2020-11-271-6/+14
| | | | | | | * do not construct DenseTensor instances directly * instead, use TensorSpec::from_expr to make a spec * and factory.from_spec() to make a value * also, use value codec directly to encode a value
* use factory.from_spec() instead of tensor::makeTensor<T>Arne Juul2020-11-271-29/+29
|
* cosmetic cleanupsArne Juul2020-10-152-4/+2
|
* simplifyArne Juul2020-10-152-6/+2
|