Control parameter feedback protocol for adapting to data stream response feedback
A method and related systems for providing adaptive responses by varying parameters used to select content includes obtaining an accuracy feedback indicator associated with a first generated message and determine an updated context generation component by applying a feedback-based protocol to the first state to set the control parameters to a second state based on the accuracy of the first generated message. Some embodiments may then generate a content sequence using the modified control parameters and use the content sequence as an input context for a second query to a language model.
1. A method for providing adaptive responses to data stream data comprising:
obtaining a first vector associated with a text sequence entered into a first client device and a second vector associated with data stream data provided by a second client device;
selecting a first vector subset by determining a ranking of first distances between content vectors and at least one of the first vector or the second vector using control parameters configured to a first state;
wherein determining the ranking of the first distances comprises determining the first distances, and wherein determining the first distances comprises determining distances between content vectors and the second vector associated with the data stream data;
determining first content to be used as input context for a language model based on the first vector subset;
obtaining a first message by sending, to the language model, a prompt comprising the text sequence and the first input context, the first input context comprising the data stream data and first content associated with the first vector subset;
updating the control parameters to a second state by using a feedback-based protocol based on the first state and a feedback indicator;
determining a second vector subset by performing search using the updated control parameters;
selecting a second vector subset by determining a ranking of second distances between the content vectors and at least one of the first vector and the second vector using the control parameters after updating the control parameters to the second state; and
obtaining a second message by sending, to the language model, the data stream data, second content mapped to the second vector subset, and the text sequence.
2. The method of claim 1 , wherein:
the feedback-based protocol comprises a neural network,
using the feedback-based protocol comprises providing, as inputs to the neural network, a parameter set and the feedback indicator, and
the parameter set comprises a parameter controlling a limit on a count of nearest neighbors, a parameter controlling a number of clusters to search, a parameter controlling a search radius, or a parameter controlling a search time.
3. The method of claim 1 , wherein the data stream data comprises voice data, and wherein obtaining the second vector comprises generating tokens based on the voice data.
4. The method of claim 1 , wherein the data stream data comprises visual data, and wherein obtaining the second vector comprises:
applying a convolutional neural network to the visual data to recognize a set of object categories; and
generating the second vector based on the set of object categories.
5. The method of claim 1 , further comprising determining a domain category associated with the text sequence, wherein determining the first distances between the content vectors and the second vector comprises selecting a set of databases comprising the content vectors based on the domain category.
6. The method of claim 5 , wherein determining the domain category comprises determining the domain category by determining domain space vector distances between the text sequence and a set of domain category vectors representing different domain categories.
7. The method of claim 1 , wherein the feedback indicator indicates a failed outcome or a negative reward, and wherein applying the feedback-based protocol comprises:
determining a set of parameter space directions of a previous update to the control parameters;
selecting an opposite direction of the set of parameter space directions; and
setting a parameter value of the second state based on the opposite direction.
8. The method of claim 1 , further comprising obtaining the feedback indicator, wherein obtaining the feedback indicator comprises:
determining a result indicating that the first message comprises an output value that exceeds a set of numeric boundaries; and
generating the feedback indicator based on the result, wherein the feedback indicator indicates a failed outcome.
9. The method of claim 1 , further comprising obtaining the feedback indicator, wherein obtaining the feedback indicator comprises:
retrieving a set of target tokens based on the text sequence or the data stream data;
determining a count of matches between the set of target tokens and the first message;
determining a result indicating that the count of matches does not satisfy a threshold; and
generating the feedback indicator based on the result, wherein the feedback indicator indicates a failed outcome.
10. One or more non-transitory, machine-readable media storing program code that, when executed by one or more processors, causes the one or more processors to perform operations for providing adaptive responses to data stream data comprising:
obtaining a first vector associated with a text sequence and a second vector associated with data stream data;
selecting a first vector subset by determining a ranking of first distances between content vectors and at least one of the first vector or the second vector using control parameters configured to a first state;
wherein determining the ranking of the first distances comprises determining the first distances, and wherein determining the first distances comprises determining distances between content vectors and the second vector associated with the data stream data;
determining first content to be used as input context for a language model based on the first vector subset;
obtaining a first message by sending, to the language model, a prompt comprising the text sequence and the first input context, the first input context comprising the data stream data and first content associated with the first vector subset;
updating the control parameters to a second state by using a feedback-based protocol based on the first state and a feedback indicator;
determining a second vector subset by performing search using the updated control parameters;
selecting a second vector subset by determining a ranking of second distances between the content vectors and at least one of the first vector and the second vector using the control parameters after updating the control parameters to the second state; and
obtaining a second message by sending, to the language model, the data stream data, second content mapped to the second vector subset, and the text sequence.
11. The one or more non-transitory, machine-readable media of claim 10 , wherein updating the control parameters to the second state comprises:
obtaining a set of tolerance boundaries;
determining a result indicating that a candidate new state satisfies the set of tolerance boundaries; and
setting the control parameters to the candidate new state based on the result.
12. The one or more non-transitory, machine-readable media of claim 10 , wherein updating the control parameters comprises modifying a parameter controlling an order of an input context.
13. The one or more non-transitory, machine-readable media of claim 10 , wherein determining the ranking of the first distances comprises:
determining the first distances between the content vectors and at least one of the first vector or the second vector using the control parameters; and
ranking the first distances.
14. The one or more non-transitory, machine-readable media of claim 10 , wherein:
the feedback-based protocol comprises a neural network,
using the feedback-based protocol comprises providing, as inputs to the neural network, a parameter set and the feedback indicator, and
the parameter set comprises a parameter controlling a limit on a count of nearest neighbors, a parameter controlling a number of clusters to search, a parameter controlling a search radius, or a parameter controlling a search time.
15. The one or more non-transitory, machine-readable media of claim 10 , the operations further comprising:
obtaining text data in an image;
applying a convolutional neural network the image to obtain text data;
generating the second vector based on the text data.
16. The one or more non-transitory, machine-readable media of claim 10 , the operations further comprising determining a domain category associated with the text sequence, wherein determining the first distances between the content vectors and the second vector comprises selecting a set of databases comprising the content vectors based on the domain category.
17. The one or more non-transitory, machine-readable media of claim 10 , the operations further comprising obtaining the feedback indicator, wherein obtaining the feedback indicator comprises:
determining a result indicating whether a count of matches between a set of target tokens associated with the data stream data and the first message satisfies a set of criteria; and
generating the feedback indicator based on the result.