Examplify

STILL WIP

Examplify is an offline CPU-first low-resource chat application to perform Retrieval-Augmented Generation (RAG) on your corpus of data. It utilises an 8-bit quantised openchat-3.6 model, running on CTranslate2's inference engine for maximum CPU performance.

Requirements

• Docker Compose
• 10 GB RAM

Benchmarks

Model	Tokens	Time (s)	Throughput (t/s)	Device
zephyr-7b-beta-ct2-int8	219	2.272	96.396	NVIDIA RTX 3090
zephyr-7b-beta-ct2-int8	211	24.482	8.619	Intel i7-8700
openchat-3.5-ct2-int8	151	0.832	181.469	NVIDIA RTX 3090
openchat-3.5-ct2-int8	156	1.573	99.160	NVIDIA RTX 3080 Ti
openchat-3.5-ct2-int8	152	10.611	14.325	Intel i7-12800H
openchat-3.5-ct2-int8	151	9.696	15.574	Intel i7-8700
openchat-3.5-ct2-int8	151	9.667	15.620	Intel i7-1260P
openchat-3.5-ct2-int8	151	20.794	7.262	Intel i9-11900H
openchat-3.6-ct2-int8	174	1.340	129.828	NVIDIA RTX 3090
openchat-3.6-ct2-int8	189	22.500	8.400	Intel i7-8700

Setup

To setup the application, we must populate your .env file. You can do this with the following.

Important

OMP_NUM_THREADS should correspond to the number of physical cores available.

{
  echo BACKEND_URL=localhost
  echo BACKEND_PORT=443
  echo CT2_USE_EXPERIMENTAL_PACKED_GEMM=1
  echo OMP_NUM_THREADS=8
} > .env

Usage

You can start the application and access the Swagger UI at https://localhost/api/schema/swagger.

Warning

Before offline usage, you must run the application at least once with internet access to install any necessary dependencies.

make u

Development

Install all dependencies with the following.

poetry install

Delete cached models.

sudo make clean