[Workbench] Bring back model config parametrization
Follow-up on a discord request to bring back full control of local model parameters as an advanced tool (e.g. temp, topK, topP, …)
QVAC | Product Team 16 days ago
QVAC/ Workbench
[Workbench] Bring back model config parametrization
Follow-up on a discord request to bring back full control of local model parameters as an advanced tool (e.g. temp, topK, topP, …)
QVAC | Product Team 16 days ago
QVAC/ Workbench
Completed
Hello, I’m reaching out because we are currently evaluating QVAC as a possible foundation for a privacy-first health AI product, and we would like to better understand the technical scope of what is available today. We are not looking only for a standalone consumer health app. We are building our own health platform with a fully custom methodology, proprietary interpretation framework, and our own structured dataset. Our stack includes: - a proprietary knowledge base, - custom health reasoning logic, - internal rules and axioms, - curated Q&A datasets, - domain-specific interpretation workflows, - and user-specific health context that we want the system to process in a controlled and consistent way. In other words, our interest is not just in using AI inside a health app, but in building our own health intelligence layer on top of a local/private AI stack. Because of that, we would like to clarify the following: 1. Is QVAC Health available only as a standalone end-user application, or do you also offer it as an SDK, white-label framework, embeddable product layer, or reusable application foundation for third-party teams? 2. If QVAC Health itself is not exposed as an SDK, can developers still build a fully custom branded health application on top of QVAC while reusing any of the underlying health-related infrastructure, data flows, local AI capabilities, or app components? 3. Does the QVAC SDK support integration of custom models, custom adapters, or fine-tuned domain-specific reasoning layers? 4. Can QVAC be used with a proprietary dataset and a custom interpretation framework, so that the resulting application follows our own health logic rather than a generic wellness model? 5. Is there support for combining: - local on-device AI, - a proprietary structured knowledge base, - user-specific health data, - and custom application logic within one product architecture? 6. Do you provide developer documentation, architectural guidance, or implementation examples for teams building their own health product on top of QVAC SDK? 7. Is it possible to maintain a fully custom data model, custom workflows, and custom UX while still relying on QVAC as the underlying local AI platform? 8. More specifically, can QVAC support a setup where the application uses: - our own branded frontend, - our own health ontology, - our own reasoning rules, - our own user-state model, - and our own proprietary dataset while keeping processing local and privacy-preserving? From our perspective, the key question is whether QVAC should be understood as: - a standalone health app for end users, or - a developer platform that can power a fully custom health AI application with third-party domain intelligence. If useful, we would be happy to share more detail about our intended architecture and the type of dataset and reasoning framework we are working with. Thank you in advance. We would appreciate any technical clarification you can provide. Best regards, Filip
filipdusek@me.com 20 days ago
Inquiries / Doubts
Completed
Hello, I’m reaching out because we are currently evaluating QVAC as a possible foundation for a privacy-first health AI product, and we would like to better understand the technical scope of what is available today. We are not looking only for a standalone consumer health app. We are building our own health platform with a fully custom methodology, proprietary interpretation framework, and our own structured dataset. Our stack includes: - a proprietary knowledge base, - custom health reasoning logic, - internal rules and axioms, - curated Q&A datasets, - domain-specific interpretation workflows, - and user-specific health context that we want the system to process in a controlled and consistent way. In other words, our interest is not just in using AI inside a health app, but in building our own health intelligence layer on top of a local/private AI stack. Because of that, we would like to clarify the following: 1. Is QVAC Health available only as a standalone end-user application, or do you also offer it as an SDK, white-label framework, embeddable product layer, or reusable application foundation for third-party teams? 2. If QVAC Health itself is not exposed as an SDK, can developers still build a fully custom branded health application on top of QVAC while reusing any of the underlying health-related infrastructure, data flows, local AI capabilities, or app components? 3. Does the QVAC SDK support integration of custom models, custom adapters, or fine-tuned domain-specific reasoning layers? 4. Can QVAC be used with a proprietary dataset and a custom interpretation framework, so that the resulting application follows our own health logic rather than a generic wellness model? 5. Is there support for combining: - local on-device AI, - a proprietary structured knowledge base, - user-specific health data, - and custom application logic within one product architecture? 6. Do you provide developer documentation, architectural guidance, or implementation examples for teams building their own health product on top of QVAC SDK? 7. Is it possible to maintain a fully custom data model, custom workflows, and custom UX while still relying on QVAC as the underlying local AI platform? 8. More specifically, can QVAC support a setup where the application uses: - our own branded frontend, - our own health ontology, - our own reasoning rules, - our own user-state model, - and our own proprietary dataset while keeping processing local and privacy-preserving? From our perspective, the key question is whether QVAC should be understood as: - a standalone health app for end users, or - a developer platform that can power a fully custom health AI application with third-party domain intelligence. If useful, we would be happy to share more detail about our intended architecture and the type of dataset and reasoning framework we are working with. Thank you in advance. We would appreciate any technical clarification you can provide. Best regards, Filip
filipdusek@me.com 20 days ago
Inquiries / Doubts
Prevent model download when using delegated inference
[from Discord] The mobile app is trying to download models every time the app reloads. I need to re-toggle the delegation enabled button on every reload. It should enable delegate modes across reloads and restarts once I have set it up. And it should not try to always download models once delegation mode is setup.
QVAC | Product Team 4 months ago
QVAC/ Workbench
Prevent model download when using delegated inference
[from Discord] The mobile app is trying to download models every time the app reloads. I need to re-toggle the delegation enabled button on every reload. It should enable delegate modes across reloads and restarts once I have set it up. And it should not try to always download models once delegation mode is setup.
QVAC | Product Team 4 months ago
QVAC/ Workbench
Completed
Hello QVAC team. I'm on the waitlist for your SDK, and have a series of questions re: business alignments. Can we arrange a meeting with someone who can speak to both the business and technical aspects?
Heather Flannery 4 months ago
Inquiries / Doubts
Completed
Hello QVAC team. I'm on the waitlist for your SDK, and have a series of questions re: business alignments. Can we arrange a meeting with someone who can speak to both the business and technical aspects?
Heather Flannery 4 months ago
Inquiries / Doubts
Completed
Support custom model import also on Desktop version
[from keet] The desktop version should support the import of gguf models straight from the URL just like the mobile versions do
QVAC | Product Team 5 months ago
Prioritized
QVAC/ Workbench
Completed
Support custom model import also on Desktop version
[from keet] The desktop version should support the import of gguf models straight from the URL just like the mobile versions do
QVAC | Product Team 5 months ago
Prioritized
QVAC/ Workbench
Completed
Registry Files Not Cleaned Up on Uninstall - PC
It’s very interesting. Anyway, was messing around with the Delegated Inference to phone on PC and it mashed everything up. Had to Uninstall and Reinstall - but the lack of a clean uninstall is created file conflicts with the old registry entries, which probably should not be there anyway.
John Lillywhite 5 months ago
QVAC/ Workbench
Completed
Registry Files Not Cleaned Up on Uninstall - PC
It’s very interesting. Anyway, was messing around with the Delegated Inference to phone on PC and it mashed everything up. Had to Uninstall and Reinstall - but the lack of a clean uninstall is created file conflicts with the old registry entries, which probably should not be there anyway.
John Lillywhite 5 months ago
QVAC/ Workbench
Online learning llm
The flow llm generate text user does some updates or corrections corrected text is used as finetuning data Next time similar prompt appears, LLM uses learned patterns and user needs to make less corrections over time
Olya Sirkin 6 months ago
Feature Requests
Online learning llm
The flow llm generate text user does some updates or corrections corrected text is used as finetuning data Next time similar prompt appears, LLM uses learned patterns and user needs to make less corrections over time
Olya Sirkin 6 months ago
Feature Requests
Multi-thread memory
[extracted from the keet communityt room] Ensure all conversations/threads are embedded so that they are accessible via RAG from other threads, instead of isolated - like a kind of “memory”. Using memory or not could be optional per thread.
QVAC | Product Team 6 months ago
Harnessing feedback
QVAC/ Workbench
Multi-thread memory
[extracted from the keet communityt room] Ensure all conversations/threads are embedded so that they are accessible via RAG from other threads, instead of isolated - like a kind of “memory”. Using memory or not could be optional per thread.
QVAC | Product Team 6 months ago
Harnessing feedback
QVAC/ Workbench
Workbench support for Apple Intel
[retrieved from the Keet community] You should add support for Apple Intel to QVAC/Workbench
Hugo 6 months ago
Harnessing feedback
QVAC/ Workbench
Workbench support for Apple Intel
[retrieved from the Keet community] You should add support for Apple Intel to QVAC/Workbench
Hugo 6 months ago
Harnessing feedback
QVAC/ Workbench
Federated Learning for Image Generation
Derivated from https://qvacbytether.featurebase.app/p/local-image-generation Federated learning can definitely enhance the quality of diffusion models while offering a privacy-preserving alternative to centralized solutions. Unlike these solutions, where user data (prompts, generated images or seed uploads) is sent to servers for processing (and potentially retained), FL enables collaborative model training across decentralised devices without sharing raw data. It remains to be solved how good this can get, how scalable it is and what nodes have more incentive in handling this effort for the benefit of all. Universities could be an example.
QVAC | Product Team 6 months ago
Harnessing feedback
Feature Requests
Federated Learning for Image Generation
Derivated from https://qvacbytether.featurebase.app/p/local-image-generation Federated learning can definitely enhance the quality of diffusion models while offering a privacy-preserving alternative to centralized solutions. Unlike these solutions, where user data (prompts, generated images or seed uploads) is sent to servers for processing (and potentially retained), FL enables collaborative model training across decentralised devices without sharing raw data. It remains to be solved how good this can get, how scalable it is and what nodes have more incentive in handling this effort for the benefit of all. Universities could be an example.
QVAC | Product Team 6 months ago
Harnessing feedback
Feature Requests
Local Image Generation
Introducing a way for people to create images locally, easily and cross-platform seems really useful for any number of applications. Cloud services expose us to risks that something like this can eliminate: exposure: there are default public galleries and no opt-out for training on some providers, which means the images are shared and used to improve models. server retention: aside from caching upload (even sensitive ones) aside from promises that data is stored in segregated locations, “deleted” content is not clear and while its claimed to not be used for training, data retention policies change all the time. Without verification, the current policies basically equate to indefinite storage. costs: generating locally is not free (has some costs) but for scale (like design schools, etc.) it’s certainly cheaper Federated learning can complement this to ensure that the models keep improving for all - based on everyone’s information - but without all the implications above.
Hugo 6 months ago
Harnessing feedback
Feature Requests
Local Image Generation
Introducing a way for people to create images locally, easily and cross-platform seems really useful for any number of applications. Cloud services expose us to risks that something like this can eliminate: exposure: there are default public galleries and no opt-out for training on some providers, which means the images are shared and used to improve models. server retention: aside from caching upload (even sensitive ones) aside from promises that data is stored in segregated locations, “deleted” content is not clear and while its claimed to not be used for training, data retention policies change all the time. Without verification, the current policies basically equate to indefinite storage. costs: generating locally is not free (has some costs) but for scale (like design schools, etc.) it’s certainly cheaper Federated learning can complement this to ensure that the models keep improving for all - based on everyone’s information - but without all the implications above.
Hugo 6 months ago
Harnessing feedback
Feature Requests