AI responses by comparison
The present technology enhances AI responses by leveraging multiple natural language processing models. The system receives an input prompt, which is processed by various Component AIs, such as ChatGPT, Google Gemini, and others. Each Component AI generates a response, which is then compared by an Ensemble AI. The Ensemble AI evaluates agreements, disagreements, and unique contributions among the responses using techniques like cosine similarity and Euclidean distance. Based on these comparisons, the Ensemble AI produces a comprehensive and reliable output. The system can dynamically select and update Component AIs based on the input's subject matter and prompt type, ensuring the most relevant models are utilized. Additionally, the system includes a fact-checking mode to resolve discrepancies and enhance response accuracy. This method ensures a more accurate and complete AI-generated response by integrating and comparing outputs from multiple AI models.
1 . A non-transitory, computer-readable storage medium for improving artificial intelligence (AI) system output comprising instructions recorded thereon, wherein the instructions, when executed by at least one data processor of a system, cause the system to:
obtain an input,
wherein the input is a natural language prompt, a natural language query, or both;
submit the input to a first component AI, a second component AI, and a third component AI,
wherein the first component AI returns a first natural language response to the input;
wherein the second component AI returns a second natural language response to the input;
wherein the third component AI returns a third natural language response to the input;
wherein the first component AI, the second component AI, and the third component AI are different component AIs;
obtain a first embedding vector from the first component AI, a second embedding vector from the second component AI, and a third embedding vector from the third component AI,
wherein the first embedding vector corresponds to a first simple sentence of the first natural language response;
wherein the second embedding vector corresponds to a second simple sentence of the second natural language response;
wherein the third embedding vector corresponds to a third simple sentence of the third natural language response;
wherein the first embedding vector, the second embedding vector, and the third embedding vector are numerical vectors in a multidimensional space;
compare the first embedding vector to the second embedding vector to detect a first agreement,
wherein detecting the first agreement comprises comparing a first distance between the first embedding vector and the second embedding vector to a first threshold distance;
in response to failing to detect the first agreement between the first embedding vector and the second embedding vector, determine whether to present the first simple sentence or the second simple sentence to a user by comparing the first embedding vector and the third embedding vector to detect a second agreement or a second disagreement,
wherein detecting the second agreement comprises comparing a second distance between the first embedding vector and the third embedding vector to the first threshold distance;
wherein detecting a first disagreement comprises comparing a third distance between the first embedding vector and the third embedding vector to a second threshold distance; and
present, upon detecting the second agreement, to the user, an output natural language response,
wherein the output natural language response is based on the first simple sentence, the third simple sentence, or both, and
wherein the output natural language response represents an improved AI system output as a result of basin the output natural language response on the first simple sentence, the third simple sentence, or both upon detecting the second agreement.
2 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
compare the first embedding vector to the second embedding vector to detect a first unique contribution,
wherein detecting the first unique contribution comprises comparing a fourth distance between the first embedding vector and the second embedding vector to a third threshold distance;
in response to detecting the first unique contribution, compare the second embedding vector and the third embedding vector to detect a second unique contribution,
wherein detecting the second unique contribution comprises comparing a fifth distance between the third embedding vector and the second embedding vector to the third threshold distance; and
present, upon detecting the first and second unique contribution, to the user, the output natural language response,
wherein the output natural language response is based on the second simple sentence.
3 . The non-transitory, computer-readable storage medium of claim 2 , comprising instructions to:
detect the first unique contribution in response to detecting that the fourth distance is approximately equal to the third threshold distance,
wherein the fourth distance is a fifth cosine similarity value between the first embedding vector and the second embedding vector;
wherein the third threshold distance is substantially similar to 0; and
detect the second unique contribution in response to detecting that the fifth distance is approximately equal to the third threshold distance,
wherein the fifth distance is a sixth cosine similarity value between the second embedding vector and the third embedding vector;
wherein the third threshold distance is substantially similar to 0.
4 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
detect the first agreement in response to detecting that the first distance is approximately equal to the first threshold distance,
wherein the first distance is a first cosine similarity value between the first embedding vector and the second embedding vector;
wherein the first threshold distance is substantially similar to 1;
detect the second agreement in response to detecting that the second distance is approximately equal to the first threshold distance,
wherein the second distance is a third cosine similarity value between the first embedding vector and the third embedding vector;
wherein the first threshold distance is substantially similar to 1;
detect the first disagreement in response to detecting that the third distance is approximately equal to the second threshold distance,
wherein the third distance is a fourth cosine similarity value between the first embedding vector and the third embedding vector;
wherein the second threshold distance is substantially similar to −1; and
detect a second disagreement in response to detecting that a sixth distance is approximately equal to the second threshold distance,
wherein the sixth distance is a second cosine similarity value between the first embedding vector and the second embedding vector;
wherein the second threshold distance is substantially similar to −1.
5 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
present, upon detecting the second agreement, to the user, the output natural language response,
wherein the output natural language response excludes the second simple sentence.
6 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
submit the input to the first component AI, the second component AI, and the third component AI,
wherein the first component AI, the second component AI, and the third component AI are pre-selected by the user.
7 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
determine a subject matter, a prompt type, or a query type of the input; and
select the first component AI, the second component AI, and the third component AI based on the subject matter, the prompt type, or the query type of the input.
8 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
record the second disagreement,
wherein recording the second disagreement increases a second disagreement total;
compare the second disagreement total to a disagreement threshold,
wherein the disagreement threshold is a predetermined integer; and
in response to detecting that the second disagreement total exceeds the disagreement threshold, replace the third component AI with a fourth component AI.
9 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
present the output natural language response,
wherein the output natural language response is structured as a list, a paragraph, or both.
10 . The non-transitory, computer-readable storage medium of claim 1 , comprising instructions to:
present the output natural language response on a virtual reality, augmented reality, or mixed reality device; and
navigate the output natural language response via a wand device associated with the virtual reality, augmented reality, or mixed reality device.
11 . A method for improving artificial intelligence (AI) system output, the method comprising:
obtaining an input,
wherein the input is a natural language prompt, a natural language query, or both;
submitting the input to a first component AI, a second component AI, and a third component AI,
wherein the first component AI returns a first natural language response to the input;
wherein the second component AI returns a second natural language response to the input;
wherein the third component AI returns a third natural language response to the input;
wherein the first component AI, the second component AI, and the third component AI are different component AIs;
obtaining a first embedding vector from the first component AI, a second embedding vector from the second component AI, and a third embedding vector from the third component AI,
wherein the first embedding vector corresponds to a first simple sentence of the first natural language response;
wherein the second embedding vector corresponds to a second simple sentence of the second natural language response;
wherein the third embedding vector corresponds to a third simple sentence of the third natural language response;
wherein the first embedding vector, the second embedding vector, and the third embedding vector are numerical vectors in a multidimensional space;
comparing the first embedding vector to the second embedding vector to detect a first agreement,
wherein detecting the first agreement comprises comparing a first distance between the first embedding vector and the second embedding vector to a first threshold distance;
in response to failing to detect the first agreement between the first embedding vector and the second embedding vector, determining whether to present the first simple sentence or the second simple sentence to a user by comparing the first embedding vector and the third embedding vector to detect a second agreement or a second disagreement,
wherein detecting the second agreement comprises comparing a second distance between the first embedding vector and the third embedding vector to the first threshold distance;
wherein detecting a first disagreement comprises comparing a third distance between the first embedding vector and the third embedding vector to a second threshold distance; and
presenting, upon detecting the second agreement, to the user, an output natural language response,
wherein the output natural language response is based on the first simple sentence, the third simple sentence, or both, and
wherein the output natural language response represents an improved AI system output as a result of basing the output natural language response on the first simple sentence, the third simple sentence, or both upon detecting the second agreement.
12 . The method of claim 11 , further comprising:
comparing the first embedding vector to the second embedding vector to detect a first unique contribution,
wherein detecting the first unique contribution comprises comparing a fourth distance between the first embedding vector and the second embedding vector to a third threshold distance;
in response to detecting the first unique contribution, comparing the second embedding vector and the third embedding vector to detect a second unique contribution,
wherein detecting the second unique contribution comprises comparing a fifth distance between the third embedding vector and the second embedding vector to the third threshold distance; and
presenting, upon detecting the first and second unique contribution, to the user, the output natural language response,
wherein the output natural language response is based on the second simple sentence.
13 . The method of claim 12 , further comprising:
detecting the first unique contribution in response to detecting that the fourth distance is approximately equal to the third threshold distance,
wherein the fourth distance is a fifth cosine similarity value between the first embedding vector and the second embedding vector;
wherein the third threshold distance is substantially similar to 0; and
detecting the second unique contribution in response to detecting that the fifth distance is approximately equal to the third threshold distance,
wherein the fifth distance is a sixth cosine similarity value between the second embedding vector and the third embedding vector;
wherein the third threshold distance is substantially similar to 0.
14 . The method of claim 11 , further comprising:
detecting the first agreement in response to detecting that the first distance is approximately equal to the first threshold distance,
wherein the first distance is a first cosine similarity value between the first embedding vector and the second embedding vector;
wherein the first threshold distance is substantially similar to 1;
detecting the second agreement in response to detecting that the second distance is approximately equal to the first threshold distance,
wherein the second distance is a third cosine similarity value between the first embedding vector and the third embedding vector;
wherein the first threshold distance is substantially similar to 1;
detecting the first disagreement in response to detecting that the third distance is approximately equal to the second threshold distance,
wherein the third distance is a fourth cosine similarity value between the first embedding vector and the third embedding vector;
wherein the second threshold distance is substantially similar to −1; and
detecting a second disagreement in response to detecting that a sixth distance is approximately equal to the second threshold distance,
wherein the sixth distance is a second cosine similarity value between the first embedding vector and the second embedding vector;
wherein the second threshold distance is substantially similar to −1.
15 . The method of claim 11 , further comprising:
presenting, upon detecting the second agreement, to the user, the output natural language response,
wherein the output natural language response excludes the second simple sentence.
16 . The method of claim 11 , further comprising:
submitting the input to the first component AI, the second component AI, and the third component AI,
wherein the first component AI, the second component AI, and the third component AI are pre-selected by the user.
17 . The method of claim 11 , further comprising:
determining a subject matter, a prompt type, or a query type of the input; and
selecting the first component AI, the second component AI, and the third component AI based on the subject matter, the prompt type, or the query type of the input.
18 . The method of claim 11 , further comprising:
recording the second disagreement,
wherein recording the second disagreement increases a second disagreement total;
comparing the second disagreement total to a disagreement threshold,
wherein the disagreement threshold is a predetermined integer; and
in response to detecting that the second disagreement total exceeds the disagreement threshold, replacing the third component AI with a fourth component AI.
19 . The method of claim 11 , further comprising:
presenting the output natural language response,
wherein the output natural language response is structured as a list, a paragraph, or both.
20 . The method of claim 11 , further comprising:
presenting the output natural language response on a virtual reality, augmented reality, or mixed reality device; and
navigating the output natural language response via a wand device associated with the virtual reality, augmented reality, or mixed reality device.
21 . A system for improving artificial intelligence (AI) system output, the system comprising:
at least one hardware processor; and
at least one non-transitory memory storing instructions, which, when executed by the at least one hardware processor, cause the system to:
obtain an input,
wherein the input is a natural language prompt, a natural language query, or both;
submit the input to a first component AI, a second component AI, and a third component AI,
wherein the first component AI returns a first natural language response to the input;
wherein the second component AI returns a second natural language response to the input;
wherein the third component AI returns a third natural language response to the input;
wherein the first component AI, the second component AI, and the third component AI are different component AIs;
obtain a first embedding vector from the first component AI, a second embedding vector from the second component AI, and a third embedding vector from the third component AI,
wherein the first embedding vector corresponds to a first simple sentence of the first natural language response;
wherein the second embedding vector corresponds to a second simple sentence of the second natural language response;
wherein the third embedding vector corresponds to a third simple sentence of the third natural language response;
wherein the first embedding vector, the second embedding vector, and the third embedding vector are numerical vectors in a multidimensional space;
compare the first embedding vector to the second embedding vector to detect a first agreement,
wherein detecting the first agreement comprises comparing a first distance between the first embedding vector and the second embedding vector to a first threshold distance;
in response to failing to detect the first agreement between the first embedding vector and the second embedding vector, determine whether to present the first simple sentence or the second simple sentence to a user by comparing the first embedding vector and the third embedding vector to detect a second agreement or a second disagreement,
wherein detecting the second agreement comprises comparing a second distance between the first embedding vector and the third embedding vector to the first threshold distance;
wherein detecting a first disagreement comprises comparing a third distance between the first embedding vector and the third embedding vector to a second threshold distance; and
present, upon detecting the second agreement, to the user, an output natural language response,
wherein the output natural language response is based on the first simple sentence, the third simple sentence, or both, and
wherein the output natural language response represents an improved AI system output as a result of basing the output natural language response on the first simple sentence, the third simple sentence, or both upon detecting the second agreement.
22 . The system of claim 21 , comprising instructions to:
compare the first embedding vector to the second embedding vector to detect a first unique contribution,
wherein detecting the first unique contribution comprises comparing a fourth distance between the first embedding vector and the second embedding vector to a third threshold distance;
in response to detecting the first unique contribution, compare the second embedding vector and the third embedding vector to detect a second unique contribution,
wherein detecting the second unique contribution comprises comparing a fifth distance between the third embedding vector and the second embedding vector to the third threshold distance; and
present, upon detecting the first and second unique contribution, to the user, the output natural language response,
wherein the output natural language response is based on the second simple sentence.
23 . The system of claim 22 , comprising instructions to:
detect the first unique contribution in response to detecting that the fourth distance is approximately equal to the third threshold distance,
wherein the fourth distance is a fifth cosine similarity value between the first embedding vector and the second embedding vector;
wherein the third threshold distance is substantially similar to 0; and
detect the second unique contribution in response to detecting that the fifth distance is approximately equal to the third threshold distance,
wherein the fifth distance is a sixth cosine similarity value between the second embedding vector and the third embedding vector;
wherein the third threshold distance is substantially similar to 0.
24 . The system of claim 21 , comprising instructions to:
detect the first agreement in response to detecting that the first distance is approximately equal to the first threshold distance,
wherein the first distance is a first cosine similarity value between the first embedding vector and the second embedding vector;
wherein the first threshold distance is substantially similar to 1;
detect the second agreement in response to detecting that the second distance is approximately equal to the first threshold distance,
wherein the second distance is a third cosine similarity value between the first embedding vector and the third embedding vector;
wherein the first threshold distance is substantially similar to 1;
detect the first disagreement in response to detecting that the third distance is approximately equal to the second threshold distance,
wherein the third distance is a fourth cosine similarity value between the first embedding vector and the third embedding vector;
wherein the second threshold distance is substantially similar to −1; and
detect a second disagreement in response to detecting that a sixth distance is approximately equal to the second threshold distance,
wherein the sixth distance is a second cosine similarity value between the first embedding vector and the second embedding vector;
wherein the second threshold distance is substantially similar to −1.
25 . The system of claim 21 , comprising instructions to:
present, upon detecting the second agreement, to the user, the output natural language response,
wherein the output natural language response excludes the second simple sentence.
26 . The system of claim 21 , comprising instructions to:
submit the input to the first component AI, the second component AI, and the third component AI,
wherein the first component AI, the second component AI, and the third component AI are pre-selected by the user.
27 . The system of claim 21 , comprising instructions to:
determine a subject matter, a prompt type, or a query type of the input; and
select the first component AI, the second component AI, and the third component AI based on the subject matter, the prompt type, or the query type of the input.
28 . The system of claim 21 , comprising instructions to:
record the second disagreement,
wherein recording the second disagreement increases a second disagreement total;
compare the second disagreement total to a disagreement threshold,
wherein the disagreement threshold is a predetermined integer; and
in response to detecting that the second disagreement total exceeds the disagreement threshold, replace the third component AI with a fourth component AI.
29 . The system of claim 21 , comprising instructions to:
present the output natural language response,
wherein the output natural language response is structured as a list, a paragraph, or both.
30 . The system of claim 21 , comprising instructions to:
present the output natural language response on a virtual reality, augmented reality, or mixed reality device; and
navigate the output natural language response via a wand device associated with the virtual reality, augmented reality, or mixed reality device.