v0.9.0 - RAG Improvements | Model Directory | Odds and Ends
Due by July 19, 2024
0% complete
Users should be able to:
- [RAG] View annotations in UI to see sources when using RAG for "chat with your docs"
- [RAG] Upload new document types, including Word, PowerPoint, and Excel (images ignored)
- [MD] Select from LLMs available in Model Directory via UI to use with an Assistant
MUST HAVE (Non-negotiable product needs that are mandatory for the team)
- […
Users should be able to:
- [RAG] View annotations in UI to see sources when using RAG for "chat with your docs"
- [RAG] Upload new document types, including Word, PowerPoint, and Excel (images ignored)
- [MD] Select from LLMs available in Model Directory via UI to use with an Assistant
MUST HAVE (Non-negotiable product needs that are mandatory for the team)
- [RAG] UI lets users upload Word, PowerPoint, and Excel files (images are ignored)
- The RAG system chunks data appropriately according to file type
- [RAG] Messages include
file_citation
and/orfile_path
annotations - [RAG] UI shows users which document(s) were used in a RAG response
- Optional (if easy): Users can select an annotation to view the passage that was used
- Optional (if easy): Users can select an annotation to view/download the original file
- [MD] Models are bundled separately from packages (no longer "baked in") for UDS deployment
SHOULD HAVE (Important initiatives that are not vital, but add significant value)
- [RAG] File uploads are queued for initial processing/ingestion
- [MD] Deploy multiple models as separate containers
- [MD] UI lets users select from one or more LLMs in the Model Directory to use with an Assistant
- [Other] Users can request long-lived API keys available via API
- [Other]
transcriptions
andtranslations
endpoints are implemented according to OpenAI API spec
COULD HAVE (Nice to have initiatives that will have a small impact if left out)
- [RAG] Initial Set of Model Evals
- Establish a list of models to evaluate, both for LLM and Embeddings
- Create a testing dataset for RAG and question/answer/ground-truth data
- Formalize a set of metrics for evaluation
- Evaluate a subset of models and present for mission hero interpretation
- [RAG] Integrate RAG eval tools into our repository & connect with LeapfrogAI
- [RAG] Implement OCR/image-analysis to extract data from embedded images in file types listed above (e.g., PowerPoint)
- [Other] UI lets users generate long-lived API keys (dependent on above work)
WILL NOT HAVE (Initiatives that are not a priority for this specific time frame)
- UI implements workflow for transcription/translation/summarization