Providing generative answers including citations to source documents
Systems and methods include pre-processing documents in cloud storage using query embeddings, providing personalized prompts to users based on documents in cloud storage, real-time anticipation of user interest in information contained in documents in cloud storage, and providing generative answers including citation to source documents in cloud storage. The system and methods generate generative machine learning model (MLM) prompts based on document portions of documents in a cloud-based content management platform. The systems and methods use the generative MLM to generate responses to prompts, and the responses include citations to the document portions used to generate the responses in order for users to verify the responses.
1. A method, comprising:
obtaining a generative machine learning model (MLM) prompt, wherein the generative MLM prompt includes an indication of a user request to generate content based on one or more documents stored in a cloud-based content management platform;
generating, using an embedding model, a first query embedding based on the generative MLM prompt;
selecting, from a plurality of documents stored in the cloud-based content management platform, a subset of the plurality of documents, wherein the selection of the subset is based on the generative MLM prompt and the first query embedding;
inputting the generative MLM prompt and the subset of the plurality of documents into a first generative MLM; and
generating, using the first generative MLM, a response, wherein the response comprises content generated by the first generative MLM, and one or more citations to one or more documents of the subset.
2. The method of claim 1 , wherein selecting the subset of the plurality of documents comprises:
inputting the generative MLM prompt into a second generative MLM;
responsive to inputting the generative MLM prompt into the second generative MLM, generating, via the second generative MLM, one or more search terms; and
performing a keyword search on the plurality of documents based on the one or more search terms.
3. The method of claim 2 , wherein:
each document of the plurality of documents comprises metadata, wherein the metadata includes a timestamp corresponding to a last time a respective document was opened; and
performing the keyword search on the plurality of documents comprises ranking a document in the plurality of documents higher based on the timestamp of the respective document.
4. The method of claim 1 , wherein selecting the subset of the plurality of documents comprises selecting a document of the plurality of documents based on the first query embedding being within a threshold similarity from a second query embedding associated with the document, and including the selected document in the subset.
5. The method of claim 1 , wherein:
each document of the plurality of documents comprises metadata, wherein the metadata indicates whether a user has permission to open a respective document; and
selecting the subset of the plurality of documents comprises selecting a document of the plurality of documents based on the metadata, and including the selected document in the subset.
6. The method of claim 1 , wherein:
each document of the plurality of documents comprises metadata, wherein the metadata includes a timestamp corresponding to a last time a respective document was modified; and
selecting the subset of the plurality of documents comprises selecting a document of the plurality of documents based on the metadata, and including the selected document in the subset.
7. The method of claim 1 , wherein a citation of the one or more citations comprises a link to a document in the subset.
8. The method of claim 1 , wherein a citation of the one or more citations comprises a link to a document portion of a document in the subset.
9. A method, comprising:
obtaining a first generative machine learning model (MLM) prompt from a user interface, wherein the first generative MLM prompt includes an indication of a user request to generate content based on one or more documents stored on a cloud-based content management platform;
inputting the first generative MLM prompt and a plurality of documents into a first generative MLM, wherein:
the plurality of documents comprises documents stored on the cloud-based content management platform, and
each document of the plurality of documents comprises a plurality of document portions, and each document portion includes an associated first query embedding;
generating, using the first generative MLM, a first generative MLM response;
generating, via an embedding model, a second query embedding based on the first generative MLM response;
selecting a subset of the plurality of document portions, wherein the first query embedding of each document portion in the subset is within a threshold similarity from the second query embedding; and
including, in the first generative MLM response, content generated by the first generative MLM, and one or more citations to document portions of the subset of the plurality of document portions.
10. The method of claim 9 , further comprising providing, on the user interface, the first generative MLM response, wherein the first generative MLM response further includes a highlighting of text of the first generative MLM response.
11. The method of claim 9 , wherein a citation of the one or more citations includes a link to a document that includes a document portion of the subset of the plurality of document portions.
12. The method of claim 11 , further comprising, responsive to a user interaction with the link to the document that includes the document portion:
opening the document in the user interface; and
highlighting the document portion in the document.
13. The method of claim 9 , further comprising:
obtaining a second generative MLM prompt from the user interface; and
responsive to inputting the second generative MLM prompt, the subset of the plurality of documents, and the first generative MLM response into the first generative MLM, generating, using the first generative MLM, a second response.
14. The method of claim 9 , wherein a document portion of the plurality of document portions comprises at least one of:
a sentence;
a paragraph; or
a section.
15. The method of claim 9 , wherein the generative MLM prompt comprises a generative MLM prompt automatically generated by a subsystem of the cloud-based content management platform.
16. A system, comprising:
a memory; and
one or more processing devices, coupled to the memory, configured to perform operations comprising:
generating, using an embedding model, a first query embedding based on a generative machine learning model (MLM) prompt, wherein the generative MLM prompt includes an indication of a user request to generate content based on one or more documents stored in a cloud-based content management platform;
selecting, from a plurality of documents stored in the cloud-based content management platform, a first subset of the plurality of documents, wherein
each document in the first subset includes an associated second query embedding, and
the selection of the first subset is based on the first query embedding being within a threshold similarity from the second query embedding associated with a respective document of the plurality of documents;
selecting, from the first subset, a second subset of documents, wherein the selection of the second subset is based on an access permission of each document in the first subset;
inputting the generative MLM prompt and the second subset of the plurality of documents into a generative MLM; and
generating, using the generative MLM, a response, wherein the response comprises content generated by the generative MLM.
17. The system of claim 16 , wherein:
each document of the second subset comprises a plurality of document portions, and each document portion includes an associated third query embedding;
the processing device is further configured to select, from the plurality of document portions of the documents of the second subset, a subset of document portions, wherein the selection of the subset of document portions is based on the first query embedding being within a first threshold similarity from a third query embedding of a plurality of third query embeddings; and
inputting the generative MLM prompt and the second subset into the generative MLM comprises inputting the generative MLM prompt and the subset of document portions into the generative MLM.
18. The system of claim 17 , wherein the operations further comprise:
generating, via the embedding model, a fourth query embedding based on the generative MLM response;
determining whether the fourth query embedding is within a second threshold similarity from a third query embedding of plurality of third query embeddings; and
responsive to determining that the fourth query embedding is not within the second threshold similarity from any third query embedding of the plurality of third query embeddings, rejecting the generative MLM response.
19. The system of claim 16 , wherein the generative MLM comprises a transformer-based large language model (LLM).
20. The system of claim 16 , wherein each of the first query embedding and the second query embeddings comprises a vector of floats.