0.9.0 - RAG support in the UI and in Apps

RAG support in Helix

We now support RAG in Helix. You can upload documents and perform RAG over them from the homepage:

We have also switched "inference" and "finetune" to the more generic and user friendly "chat" and "learn":

The default Learn mode is now RAG, because it's much, much faster than fine-tuning. RAG is better at retrieving specific facts, whereas fine-tuning is better at answering general questions about the documents uploaded.

You can still fine-tune, either choose fine tuning from the app homepage, or use the settings button:

Using RAG and fine tuned data sources in Helix Apps

You can now also specify RAG and finetune data sources in Helix Apps' helix.yaml to customize an assistant with a RAG data source or fine tuned LLM. To do this, run a RAG or finetune session which will now create a "data source ID". Retrieve the rag_source_data_entity_id from the info button in a RAG session, like this:

        "rag_source_data_entity_id": "c6cc22d3-23a6-4b2d-acdd-6f561158e0c0",

And place it in a helix.yaml file in a GitHub repo like this:

name: My Test Helix RAG App
description: This is a test Helix RAG app
assistants:
- name: My Example RAG Assistant
  description: This is an example assistant with a rag source
  rag_source_id: 8b4ff837-b42e-41d2-a5cd-fc7f6c26e08f

Then proceed to use Helix Apps as documented here

This rag_source_id can also be overriden as an API parameter when making an API call.

You can do the same with finetune data sources, named finetune_data_entity_id in the info panel and specified in the helix.yaml as lora_id.

PRs in this release

• Fix/bump msg limit by @rusenask in #303
• fix(llamaindex): use text for filename in database to accommodate long filenames by @philwinder in #304
• Feature/basic data entities by @lukemarsden in #300

Full Changelog: 0.8.7...0.9.0

0.8.7 - faster, more compatible docker images with pre-baked weights

We are now publishing runner images with a variety of pre-pulled models:

• registry.helix.ml/helix/runner:0.8.7 - no pre-pulled models
• registry.helix.ml/helix/runner:0.8.7-small = llama3:instruct;phi3:instruct
• registry.helix.ml/helix/runner:0.8.7-large
  • base layer = llama3:instruct;llama3:70b;mixtral:instruct;phi3:instruct
  • upper layer = adrienbrault/nous-hermes2theta-llama3-8b:q8_0
• these images were previously zstd compressed, but that lead to instability in the build pipeline and incompatibility with openshift, so here we revert back to gzip-0 compression which is at least fast and has good compatibility

---

Full Changelog: 0.8.6...0.8.7

0.8.6 - new registry with cached model weights

What's Changed

We are now publishing runner images with a variety of pre-pulled models:

• registry.helix.ml/helix/runner:0.8.6 - no pre-pulled models
• registry.helix.ml/helix/runner:0.8.6-small - small = llama3:instruct;phi3:instruct
• registry.helix.ml/helix/runner:0.8.6-large
  • base layer = llama3:instruct;llama3:70b;mixtral:instruct;phi3:instruct
  • upper layer = adrienbrault/nous-hermes2theta-llama3-8b:q8_0

---

• two different model layers in docker by @lukemarsden in #301
• switch to new base images by @lukemarsden in #302

Full Changelog: 0.8.5...0.8.6

0.8.5 - gptscript runner, new registry

What's Changed

New gptscript container runner for private deployments. Preparation for publishing images to new registry.helix.ml

• Feature/gptscript runner by @rusenask in #295

Full Changelog: 0.8.4...0.8.5

0.8.4 - publish new docker images

What's Changed

Now publishing -small and -large docker images containing pre-baked llama3-8b & phi3-mini, and all models advertised in the UI respectively. For ollama images only.

Full Changelog: 0.8.3...0.8.4

0.8.3 - Large docker image with pre-baked weights

What's Changed

Publish a new docker image with pre-baked weights for

llama3:instruct,llama3:70b,mixtral:instruct,adrienbrault/nous-hermes2theta-llama3-8b:q8_0,phi3:instruct

Also testing https://huggingface.co/NousResearch/Hermes-2-Theta-Llama-3-8B-GGUF for JSON output, function calling & tools use.

• update models and bake a big docker image by @lukemarsden in #298

Full Changelog: 0.8.2...0.8.3

0.8.2 - Phi-3, new app homepage, UI updates, add account page

New features

Added Phi-3 model (high quality model with low memory usage), updated visual style, improved mobile views, and enabling account management page in Keycloak.

What's Changed

• Enable account mgmt by @chocobar in #291
• Feature/app homepage by @lukemarsden in #292
• add colour by @lukemarsden in #293
• changes to add phi3 model by @ObianuoObi in #289
• Feature/new create UI by @binocarlos in #294
• Fix/responsive layout by @binocarlos in #296

New Contributors

• @ObianuoObi made their first contribution in #289

Full Changelog: 0.8.1...0.8.2

0.8.1

What's Changed

• fix auth for OpenAI API by @rusenask in #290

Full Changelog: 0.8.0...0.8.1

0.8.0 - App tools & internal qapair gen

What's Changed

API Tools in Helix Apps

You can now version control configuration for OpenAPI tools integrations and spawn embedded widget sessions from such apps. You can also use the chat completion API with the new per-app API key to get tools enabled on that chat session.

Example config: https://github.com/helixml/example-helix-app

Fork the repo above and add it as an app inside Helix. Then embed the widget in an html page, activate it, and ask it about the hiring pipeline (demo API).

https://twitter.com/lmarsden/status/1787846737870127426

Fully local qapair generation

We now also have basic support (without schema enforcement, yet) for configuring the qapair generator in fine-tuning to use fully local models.

• Feature/qa pair gen by @rusenask in #284
• Feature/app tools by @binocarlos in #286
• fix subrouter by @rusenask in #287

Full Changelog: 0.7.5...0.8.0

0.7.5 - adjust ollama3-70b memory

Llama3-70B now runs quickly in 39GB of vRAM on latest ollama build.

Full Changelog: 0.7.4...0.7.5