Home

Welcome to the mamai wiki!

Target Group

1. Midwives and nurses (definition? Applied to low risk pregnancy? What do they take care of and what they cannot take care of https://my.clevelandclinic.org/health/articles/22648-midwife)

Goals

1. On device (offline) generation
2. Quality response
3. For nurses/midwives/pregnant women
4. Swahili and English support

Things to work on

Read the documents of tech4dev to be clear of the goals and work plan

Setup

1. Replace the initial download of 4.4G model files (see issue #1)
2. Fix the GPU starting failure on my Pixel 7.
3. Support Swahili conversation

Conversation

1. Enable multi turn conversation
2. Enable chat history
3. Other chat features like ChatGPT/Claude app

Evaluation

1. Find / Create the dataset to evaluate the answer quality.
2. Evaluate the Swahili ability.
3. Evaluate generation speed.
4. Design test cases to let the real users test

Documents

1. Examine input documents quality
2. Check document segmentation strategy.
3. Make the document preprocessing easy so that people could add any document they want, good for the community use.

RAG

1. How many documents to retrieve, how to retrieve (BM25, embedding similairty)
2. Check and improve prompts for generation

Model

1. Currently we use Gemma 3n E4B IT. Other choices include Gemma 3n E2B IT, Gemma 3 1B/270M, MedGemma, or maybe other small models/medical models that could fit on device.
2. Improve Swahili ability of the model (fine-tuning, add a small translation module?)

Input modality

1. Enable multimodality input (images, video, audio)
2. For audio input, we use Whisper for text transcription?

Avoid user jail breaking with prompt engineering. Do the basic protection. Do not answer irrelevant questions.

Refer to the Gemma 3n challenge for good ideas

Other problems in _scratch/tofix.md