Validating generative artificial intelligence output
Methods, apparatuses, and computer-program products are disclosed. The method may include training a generative artificial intelligence (AI) model on a plurality of data sources and generating, based on the training, training log metadata indicating individual data sources of the plurality of data sources. The method may include receiving, from a user device, a generative AI query and generating, using the trained generative AI model and based on one or more data sources of the plurality of data sources, a response to the generative AI query. The method may include mapping one or more portions of the response to the one or more data sources of the plurality of data sources based on the training log metadata and transmitting, to the user device, the response and one or more indications of the one or more data sources based on the mapping and the training log metadata.
1 . A method for analyzing a generative artificial intelligence (AI) model, comprising:
training the generative AI model on a plurality of data sources;
generating, based at least in part on the training, training log metadata indicating individual data sources of the plurality of data sources;
receiving, from a user device, a generative AI query;
generating, using the trained generative AI model and based at least in part on one or more data sources of the plurality of data sources, a response to the generative AI query;
performing a first natural language analysis to tokenize information of the one or more data sources and generate a plurality of data source tokens;
performing a second natural language analysis to tokenize contents of the response and generate a plurality of response tokens;
mapping individual data source tokens of the plurality of data source tokens to individual response tokens of the plurality of response tokens;
mapping, based at least in part on the mapping of the individual data source tokens to the individual response tokens, one or more portions of the response to the one or more data sources of the plurality of data sources based at least in part on the training log metadata; and
transmitting, to the user device, the response and one or more indications of the one or more data sources based at least in part on the mapping and the training log metadata.
2 . The method of claim 1 , further comprising:
generating the one or more indications of the one or more data sources based at least in part on the mapping of the individual data source tokens to the individual response tokens.
3 . The method of claim 1 , further comprising:
analyzing the response to determine a confidence score associated with the response; and
determining one or more actions to be performed based at least in part on the confidence score satisfying a confidence score threshold or determining one or more actions to be performed based at least in part on the confidence score failing to satisfy the confidence score threshold.
4 . The method of claim 1 , further comprising:
receiving user input indicating the plurality of data sources; and
generating a training procedure for the generative AI model, the training procedure indicating the plurality of data sources;
wherein training the generative AI model is based at least in part on the training procedure.
5 . The method of claim 4 , further comprising:
generating the one or more indications of the one or more data sources based at least in part on the training procedure.
6 . The method of claim 1 , further comprising:
verifying, with a second generative AI model and based at least in part on the response, that the response is responsive to the generative AI query.
7 . The method of claim 1 , further comprising:
analyzing the response and the one or more data sources to determine whether the response correctly includes information included in the one or more data sources.
8 . The method of claim 1 , wherein the one or more indications of the one or more data sources comprise a data source identifier, a data source location, a web address, an excerpt from the one or more data sources, or any combination thereof.
9 . An apparatus for analyzing a generative artificial intelligence (AI) model, comprising:
one or more memories storing processor-executable code; and
one or more processors coupled with the one or more memories and individually or collectively operable to execute the code to cause the apparatus to:
train the generative AI model on a plurality of data sources;
generate, based at least in part on the training, training log metadata indicating individual data sources of the plurality of data sources;
receive, from a user device, a generative AI query;
generate, using the trained generative AI model and based at least in part on one or more data sources of the plurality of data sources, a response to the generative AI query;
perform a first natural language analysis to tokenize information of the one or more data sources and generate a plurality of data source tokens;
perform a second natural language analysis to tokenize contents of the response and generate a plurality of response tokens;
map individual data source tokens of the plurality of data source tokens to individual response tokens of the plurality of response tokens;
map, based at least in part on the mapping of the individual data source tokens to the individual response tokens, one or more portions of the response to the one or more data sources of the plurality of data sources based at least in part on the training log metadata; and
transmit, to the user device, the response and one or more indications of the one or more data sources based at least in part on the mapping and the training log metadata.
10 . The apparatus of claim 9 , wherein the one or more processors are individually or collectively further operable to execute the code to cause the apparatus to:
generate the one or more indications of the one or more data sources based at least in part on the mapping of the individual data source tokens to the individual response tokens.
11 . The apparatus of claim 9 , wherein the one or more processors are individually or collectively further operable to execute the code to cause the apparatus to:
analyze the response to determine a confidence score associated with the response; and
determine one or more actions to be performed based at least in part on the confidence score satisfying a confidence score threshold or determining one or more actions to be performed based at least in part on the confidence score failing to satisfy the confidence score threshold.
12 . The apparatus of claim 9 , wherein the one or more processors are individually or collectively further operable to execute the code to cause the apparatus to:
receive user input indicating the plurality of data sources; and
generate a training procedure for the generative AI model, the training procedure indicating the plurality of data sources;
wherein training the generative AI model is based at least in part on the training procedure.
13 . The apparatus of claim 12 , wherein the one or more processors are individually or collectively further operable to execute the code to cause the apparatus to:
generate the one or more indications of the one or more data sources based at least in part on the training procedure.
14 . The apparatus of claim 9 , wherein the one or more processors are individually or collectively further operable to execute the code to cause the apparatus to:
verify, with a second generative AI model and based at least in part on the response, that the response is responsive to the generative AI query.
15 . The apparatus of claim 9 , wherein the one or more processors are individually or collectively further operable to execute the code to cause the apparatus to:
analyze the response and the one or more data sources to determine whether the response correctly includes information included in the one or more data sources.
16 . The apparatus of claim 9 , wherein the one or more indications of the one or more data sources comprise a data source identifier, a data source location, a web address, an excerpt from the one or more data sources, or any combination thereof.
17 . A non-transitory computer-readable medium storing code for analyzing a generative artificial intelligence (AI) model, the code comprising instructions executable by one or more processors to:
train the generative AI model on a plurality of data sources;
generate, based at least in part on the training, training log metadata indicating individual data sources of the plurality of data sources;
receive, from a user device, a generative AI query;
generate, using the trained generative AI model and based at least in part on one or more data sources of the plurality of data sources, a response to the generative AI query;
perform a first natural language analysis to tokenize information of the one or more data sources and generate a plurality of data source tokens;
perform a second natural language analysis to tokenize contents of the response and generate a plurality of response tokens;
map individual data source tokens of the plurality of data source tokens to individual response tokens of the plurality of response tokens;
map, based at least in part on the mapping of the individual data source tokens to the individual response tokens, one or more portions of the response to the one or more data sources of the plurality of data sources based at least in part on the training log metadata; and
transmit, to the user device, the response and one or more indications of the one or more data sources based at least in part on the mapping and the training log metadata.