Knowledge attribution and answer validation platform for generative AI output

1 . A knowledge attribution system for providing proof-of-answer for generative artificial intelligence output, comprising:

a control server having a computer storage device, operably connected to a user device;

a knowledge base stored using the computer storage device, the knowledge base comprising one or more source documents, each source document having one or more document segments;

a knowledge database stored using the computer storage device, the knowledge database comprising a plurality of knowledge elements, each knowledge element representing one of the document segments of each source document, each knowledge element having attribution metadata associating the knowledge element with the originating document segment, the attribution metadata further comprises a source attribution identifier and a source position identifier, the source attribution identifier references the source document from which the knowledge element is derived, and the source position identifier references the corresponding document segment of the knowledge element, the knowledge database is configured as a segregated database comprising a plurality of partitions, with each of the partitions containing the knowledge elements of one of the source documents, and each document segment of each source document has text content, the text content of each document segment is represented in the corresponding knowledge element as an encoded text portion, the encoded text portion corresponding to a vector representation which captures semantic and contextual meaning present within the text content in order to facilitate the similarity search which identifies the response candidate elements; and

a knowledge attribution engine executed by the control server, the knowledge attribution engine having a response generation module, the knowledge attribution engine is adapted to receive a user query entered via the user device, conduct a similarity search of the knowledge database directed at the knowledge elements of one of the source documents via the corresponding partition, and retrieve one or more response candidate elements as search results, each response candidate element corresponding to one of the knowledge elements which are relevant to the user query, the knowledge attribution engine is further adapted to generate an augmented query which instructs the response generation module to produce a response generation output utilizing at least one of the response candidate elements, wherein the augmented query includes the text content of each response candidate element, and further includes a generative restriction command instructing the response generation module to produce the generated response text only using the text content of the response candidate elements, the response generation output comprising a generated response text which answers the user query, and the response generation output further includes the attribution metadata of each response candidate element which is utilized to produce the generated response text, the knowledge attribution engine is further adapted to generate and display a a proof of answer image of each of the utilized response candidate elements using the attribution metadata within the response generation output, to serve as proof-of-answer for the generated response text, wherein the proof of answer image comprises a screenshot of each of the document segments of the source document which are identified by the attribution metadata of the response text, and wherein the generated response text and the knowledge validation image are displayed via the user device.

2 . The knowledge attribution system of claim 1 , wherein:

the knowledge attribution engine further comprises a document ingestion module, the document ingestion module is adapted to populate the knowledge database by processing each of the source documents within the knowledge base and dividing the source document into its constituent document segments, wherein each of the knowledge elements is derived from one of the document segments of each of the source documents.

3 . The knowledge attribution system of claim 1 , wherein:

each document segment represents one or more pages of the source document; and

the proof of answer image of each document segment visually reproduces non-textual content present within the document segment in addition to the text content thereof.

4 . The knowledge attribution system of claim 1 , further comprising a response presentation interface, the response presentation interface is implemented upon the user device, the response presentation interface is adapted to display the generated response text and the corresponding proof of answer image for each of the utilized response candidate elements.

5 . The knowledge attribution system of claim 1 , further comprising a knowledge attribution engine application programming interface endpoint, wherein the generated response text and the corresponding proof of answer image for each of the utilized response candidate elements are made available for external retrieval by the knowledge attribution engine via the knowledge attribution engine application programming interface endpoint.

6 . The knowledge attribution system of claim 1 , wherein the user query further identifies the source document towards which the similarity search is directed.

7 . The knowledge attribution system of claim 6 , wherein the user query is submitted to the knowledge attribution system via a knowledge attribution engine application programming interface.

8 . A method for providing proof-of-answer for generative artificial intelligence output, comprising the steps of:

providing a knowledge base comprising one or more source documents, each comprising one or more document segments containing text content;

ingesting each of the source documents to create a knowledge element for each document segment of the source document, representing the text content of each document segment as an encoded text portion within the corresponding knowledge element, using a vector representation to capture semantic and contextual meaning for the similarity search, and creating attribution metadata which associates the knowledge element with the originating document segment, the attribution metadata of each knowledge element comprising a source attribution identifier and a source position identifier, the source attribution identifier referencing the source document from which the knowledge element is derived, and the source position identifier referencing the corresponding document segment of the knowledge element;

populating a knowledge database by storing the knowledge elements within the knowledge database;

configuring the knowledge database as a segregated database with a plurality of partitions, each partition containing the knowledge elements of one of the source documents;

receiving a user query by a knowledge attribution engine;

conducting a similarity search of the knowledge database directed at the partition containing the knowledge elements of one of the source documents to retrieve response candidate elements, each response candidate element corresponding to one of the knowledge elements which is relevant to the user query;

generating an augmented query that includes the text content of each response candidate element, and instructs a response generation module to use at least one of the response candidate elements to produce a response generation output, the augmented query further instructing the response generation module to produce a generated response text using only the text content of the response candidate elements;

producing the response generation output, the response generation output comprising an answer to the user query in the form of the generated response text, the response generation output further comprising the attribution metadata of each response candidate element utilized to produce the generated response text;

generating a proof of answer image for each utilized response candidate element using the attribution metadata, the proof of answer image comprising a screenshot of each of the document segments of the source document which are identified by the attribution metadata of the response text, and presenting each proof of answer image as proof-of-answer for the generated response text which visually represents the corresponding source document.

9 . The method as recited in claim 8 , wherein:

each document segment represents one or more pages of the source document; and

the step generating a proof of answer image further comprises the proof of answer image visually reproducing non-textual content present within the document segment in addition to the text content.

10 . The method as recited in claim 8 , further comprising:

implementing a response presentation interface on a user device; and

adapting the response presentation interface to display the generated response text in conjunction with the proof of knowledge image for each utilized response candidate element.

11 . The method as recited in claim 8 , further comprising:

making the generated response text and the corresponding proof of answer images available for external retrieval via a knowledge attribution engine application programming interface endpoint.

12 . The method as recited in claim 11 , wherein the user query is submitted to the knowledge attribution engine via the application programming interface endpoint.

13 . The method as recited in claim 8 , wherein the user query further identifies the source document towards which the similarity search is directed.