Memetic database
AssigneeHoltwork LLC
InventorBrian Holt
Life of this patent
Estimating the term…
Abstract
Computer-implemented techniques for representing, storing, retrieving, and reasoning over knowledge are disclosed herein. One aspect is an axis-lattice query language paradigm that turns every query string into a rectangular Matrix→Vector→Limit→Data tensor whose shape is dictated by token position. The invention improves the operations of a computing device by increasing the amount of significant information manipulated per unit of time by a processor.
Claims15
A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
- receiving an input string;
- parsing the input string into a plurality of tokens, at least one token being a relative token;
- associating each token with a coordinate tuple having at least two coordinates;
- associating the relative token with at least one referenced coordinate tuple according to the relative token's coordinate tuple and at least one predefined relative coordinate offset;
- retrieving a first value associated with the referenced coordinate tuple;
- performing a datastore operation using the first value and a comparison operator to retrieve a plurality of data entities;
- and
- outputting the plurality of data entities.
A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
- receiving an input string;
- parsing the input string into a plurality of tokens;
- associating each of at least three tokens from the plurality of tokens with a coordinate tuple comprising at least a first coordinate index and a second coordinate index;
- counting a number of distinct second coordinate indices for each first coordinate index;
- determining that a targeted first coordinate index has fewer distinct second coordinate indices than a precedent first coordinate index;
- selecting a precedent token whose first coordinate index equals the precedent first coordinate index and whose second coordinate index is absent from any coordinate tuple whose first coordinate index equals the targeted first coordinate index;
- and
- writing a carried-forward token corresponding to the precedent token and associated with a coordinate tuple comprising the targeted first coordinate index and the second coordinate index of the precedent token.
A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
- receiving an input string encoded according to an axial grammar having a first separator token and a second separator token;
- parsing the input string into a plurality of tokens and mapping the tokens into a multidimensional lattice according to positions of instances of the first separator token and the second separator token;
- retrieving data from a datastore by evaluating constraints encoded in the multidimensional lattice;
- and
- outputting an output string comprising data from the datastore encoded according to the axial grammar, wherein the output string is a valid input string.
A computer-implemented method for performing large language model retrieval-augmented generation with an axial grammar, comprising:
- receiving a natural-language input describing a desired behavior;
- providing the natural-language input to a large language model trained on text paired with axial grammar code;
- generating, by the large language model, a sequence of tokens conforming to the axial grammar, the sequence comprising at least a first separator token marking a first dimension and a second separator token marking a second dimension;
- parsing the axial grammar code into a multidimensional lattice according to a position of the first separator token and a position of the second separator token;
- retrieving data from a datastore by evaluating constraints encoded in the multidimensional lattice;
- and
- outputting an output string comprising data from the datastore.
Dependencies are inferred from the claim text.