Techniques for automated query response determination using a hybrid artificial intelligence (AI) model
Techniques for automated query response determination using a hybrid AI are disclosed herein. An example computer-implemented method includes receiving a data file and a request including at least one query associated with the data file and applying a hybrid model to the data file. Applying the hybrid model includes segmenting the data file into one or more portions, embedding the one or more portions into a vector space, extracting, from the data file, data associated with one or more classifications, storing (i) the embedded portions in a first database and (ii) the extracted data in a second database, and determining a response to request queries based on the embedded portions and the extracted data, wherein the hybrid model constrains each response based on a parametric input prompt associated with the extracted data. The example computer-implemented method further includes storing one or more data objects indicating each response.
1 . A computer-implemented method comprising:
receiving, by one or more processors, a data file and a request including a query associated with the data file;
applying, by the one or more processors, a hybrid model to the data file, wherein applying the hybrid model comprises:
segmenting the data file into one or more portions,
embedding the one or more portions into a vector space,
extracting, from the data file, data associated with one or more classifications,
storing (i) the embedded one or more portions in a first database and (ii) the extracted data in a second database,
determining, based on the extracted data, a parametric input prompt configured to prevent the hybrid model from determining a response to the query that is unsupported by the extracted data, and
determining the response to the query in the request based on the embedded one or more portions, the extracted data, and the parametric input prompt; and
storing, by the one or more processors, a data object indicating the response.
2 . The computer-implemented method of claim 1 , wherein the hybrid model includes a large language model (LLM) and a symbolic artificial intelligence (S-AI) model.
3 . The computer-implemented method of claim 2 , wherein the LLM is configured to embed the one or more portions into the vector space.
4 . The computer-implemented method of claim 2 , wherein the S-AI model is configured to extract the data from the data file.
5 . The computer-implemented method of claim 1 , wherein segmenting the data file comprises:
extracting one or more attributes from the data file; and
segmenting the data file based on one or more segmentation windows that each include one of the one or more portions.
6 . The computer-implemented method of claim 5 , wherein storing the embedded one or more portions in the first database comprises:
storing the one or more attributes in the first database with the embedded one or more portions.
7 . The computer-implemented method of claim 1 , wherein determining the response to the query in the request comprises:
determining, by a question-answering LLM (QA-LLM), one or more sub-queries corresponding to the query based on a prompt-based instruction and a curated example; and
sequentially answering each of the one or more sub-queries to determine the response to the query.
8 . The computer-implemented method of claim 1 , wherein determining the response to the query in the request comprises:
determining a first information extraction tool (IET) of a plurality of IETs to search the first database and the second database;
searching, by the first IET, the first database to retrieve (i) the embedded one or more portions stored in the first database and (ii) the extracted data stored in the second database; and
structuring, by the first IET, the response to the query based on the searching.
9 . The computer-implemented method of claim 8 , wherein each IET of the plurality of IETs utilizes a large language model (LLM) trained to perform (i) the searching and (ii) the structuring.
10 . The computer-implemented method of claim 8 , wherein each IET of the plurality of IETs includes a corresponding description, and determining the first IET comprises:
determining a vector match similarity value between embeddings of the query and the corresponding description of each IET; and
determining the first IET for the query based on the first IET having a highest one of the vector match similarity value of each IET.
11 . The computer-implemented method of claim 8 , wherein determining the response to the query in the request comprises:
verifying, by a fusion LLM (F-LLM), the response from the first IET based on a prompt-based weighting instruction.
12 . The computer-implemented method of claim 11 , wherein the F-LLM is trained to resolve conflicts between information indicated by (i) the embedded one or more portions stored in the first database and (ii) the extracted data stored in the second database by utilizing few-shot prompting.
13 . The computer-implemented method of claim 1 , wherein storing the extracted data in the second database comprises:
storing, for each data point of the extracted data, a phrase definition in the second database, wherein each phrase definition indicates a classification of the one or more classifications associated with the data point.
14 . The computer-implemented method of claim 1 , wherein generating the data object comprises:
structuring the data object to include (i) a binary output and (ii) supporting evidence from the first database and the second database for the response.
15 . A system comprising:
one or more processors; and
at least one memory storing processor-executable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
receiving a data file and a request including a query associated with the data file;
applying a hybrid model to the data file, wherein applying the hybrid model comprises:
segmenting the data file into one or more portions,
embedding the one or more portions into a vector space,
extracting, from the data file, data associated with one or more classifications,
storing (i) the embedded one or more portions in a first database and (ii) the extracted data in a second database,
determining, based on the extracted data, a parametric input prompt configured to prevent the hybrid model from determining a response to the query that is unsupported by the extracted data, and
determining the response to the query in the request based on the embedded one or more portions, the extracted data, and the parametric input prompt; and
storing a data object indicating the response.
16 . The system of claim 15 , wherein the hybrid model includes a large language model (LLM) and a symbolic artificial intelligence (S-AI) model.
17 . The system of claim 15 , wherein determining the response to the query in the request comprises:
determining, by a question-answering LLM (QA-LLM), one or more sub-queries corresponding to the query based on a prompt-based instruction and a curated example; and
sequentially answering each of the one or more sub-queries to determine the response to the query.
18 . The system of claim 15 , wherein determining the response to the query in the request comprises:
determining a first information extraction tool (IET) of a plurality of IETs to search the first database and the second database;
searching, by the first IET, the first database to retrieve (i) the embedded one or more portions stored in the first database and (ii) the extracted data stored in the second database; and
structuring, by the first IET, the response to the query based on the searching.
19 . The system of claim 18 , wherein each IET includes a corresponding description, and determining the first IET comprises:
determining a vector match similarity value between embeddings of the query and the corresponding description of each IET; and
determining the first IET for the query based on the first IET having a highest one of the vector match similarity value of each IET.
20 . One or more non-transitory computer-readable media storing processor-executable instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
receiving a data file and a request including a query associated with the data file;
applying a hybrid model to the data file, wherein applying the hybrid model comprises:
segmenting the data file into one or more portions,
embedding the one or more portions into a vector space,
extracting, from the data file, data associated with one or more classifications,
storing (i) the embedded one or more portions in a first database and (ii) the extracted data in a second database,
determining, based on the extracted data, a parametric input prompt configured to prevent the hybrid model from determining a response to the query that is unsupported by the extracted data, and
determining the response to the query in the request based on the embedded one or more portions, the extracted data, and the parametric input prompt; and
storing a data object indicating the response.