TENSOR-NAME INDEXING
Find model tensors by their names.Find model tensors by their names.
Your model tooling knows the names, shapes and locations of its tensors. Store those descriptors in Keldra and use an exact tensor-name index to locate the corresponding stored data within a configured model. Keep the original model bytes in object storage.
Look up tensor names
Select a named tensor within the model configured for the index.
Fits your tooling
Your pipeline supplies descriptors using the model names it already understands.
Original files retained
Keep model data as objects without converting the original files to a proprietary export format.
Look up tensor names
Select a named tensor within the model configured for the index.
Fits your tooling
Your pipeline supplies descriptors using the model names it already understands.
Original files retained
Keep model data as objects without converting the original files to a proprietary export format.
Look up tensor names
Select a named tensor within the model configured for the index.
Fits your tooling
Your pipeline supplies descriptors using the model names it already understands.
Original files retained
Keep model data as objects without converting the original files to a proprietary export format.
Model files and named tensor descriptors, kept together.
How tensor-name indexing works
Store the model data, publish its tensor descriptors and configure the index for that model. A named query matches the descriptor and its source reference. This is not an automatic parser for every model-file format.
Store the model object
Keep the bytes your model tooling produces. Descriptors refer to the source path and version, with offset, length, data type and shape information.
Index the descriptor
Provide the model ID and tensor name alongside the source description. Keldra indexes the fields for the declared tensor projection.
{
"model_id": "encoder-v1",
"tensor_name": "encoder.weight",
"source_path": "models/encoder.bin",
"source_version": 0,
"offset": 0,
"length": 64,
"dtype": "f32",
"shape": [4, 4]
}
Find the stored source
An exact tensor-name query within that configured model returns matching source references. Your application can then use the relevant stored data through the supported object APIs.
Store the model object
Keep the bytes your model tooling produces. Descriptors refer to the source path and version, with offset, length, data type and shape information.
Index the descriptor
Provide the model ID and tensor name alongside the source description. Keldra indexes the fields for the declared tensor projection.
{
"model_id": "encoder-v1",
"tensor_name": "encoder.weight",
"source_path": "models/encoder.bin",
"source_version": 0,
"offset": 0,
"length": 64,
"dtype": "f32",
"shape": [4, 4]
}
Find the stored source
An exact tensor-name query within that configured model returns matching source references. Your application can then use the relevant stored data through the supported object APIs.
Locate weights without a separate tensor inventory
Address weights by their names
A model filename does not identify an individual weight. Named descriptors give model tooling an explicit lookup for the tensor it needs.
Keep the inventory with the data
Store your descriptors and model objects in the same storage platform rather than building another inventory just to track tensor locations.
Keep control of the files
Your model files remain objects on infrastructure you operate. Indexing a descriptor does not run inference, train a model or rewrite its weights.
Address weights by their names
A model filename does not identify an individual weight. Named descriptors give model tooling an explicit lookup for the tensor it needs.
Keep the inventory with the data
Store your descriptors and model objects in the same storage platform rather than building another inventory just to track tensor locations.
Keep control of the files
Your model files remain objects on infrastructure you operate. Indexing a descriptor does not run inference, train a model or rewrite its weights.
Address weights by their names
A model filename does not identify an individual weight. Named descriptors give model tooling an explicit lookup for the tensor it needs.
Keep the inventory with the data
Store your descriptors and model objects in the same storage platform rather than building another inventory just to track tensor locations.
Keep control of the files
Your model files remain objects on infrastructure you operate. Indexing a descriptor does not run inference, train a model or rewrite its weights.
When you know the weight's name
Your application needs encoder.weight from encoder-v1. The index definition selects encoder-v1; the query supplies encoder.weight. The result connects that name to the descriptor's source object rather than searching every model in the cluster.
TensorIndexSpec { model_id: "encoder-v1" }
TensorIndexQuery { tensor_name: "encoder.weight" }
A precise lookup, not a model execution engine
The tensor query selects a name within one configured model. It does not provide element-wise calculations, arbitrary tensor slicing, numerical analytics or automatic cross-model discovery. Your model tooling remains responsible for creating accurate descriptors.
Vector similarity is a different index
Use vector indexing to find items with similar embeddings. Tensor indexing instead addresses a named weight and its stored source description.
GET STARTED
Store model bytes and index their tensor descriptors.
Start Keldra, retain your original model objects and add the named descriptors your application needs.