v0.1.25

🎉 Introduction to new functions of GPTCache

1. Support the DocArray vector database

from gptcache.manager import manager_factory

data_manager = manager_factory("sqlite,docarray")

2. Add rwkv model for embedding

from gptcache.embedding import Rwkv

test_sentence = 'Hello, world.'
encoder = Rwkv(model='sgugger/rwkv-430M-pile')
embed = encoder.to_embeddings(test_sentence)

What's Changed

• [skip ci]Add workflow to publish release image by @Bennu-Li in #345
• Update the doc directory by @SimFG in #348
• Add the docker image doc by @SimFG in #349
• DocArray as a vectorstore by @jupyterjazz in #351
• Fix the doc generation failure by @SimFG in #352
• Replace base image and simplify dockerfile by @Chiiizzzy in #353
• Example with DocArray by @jupyterjazz in #354
• Add comments for session by @shiyu22 in #355
• DocArray example adjustment by @jupyterjazz in #356
• Improve the generation doc script by @SimFG in #358
• Change the model test cases as L2 cases by @SimFG in #362
• Add concat_context. by @wxywb in #365
• Add the pre/report/session docs by @SimFG in #364
• Add rwkv model for embedding. by @wxywb in #363
• Update the version to 0.1.25 by @SimFG in #367

New Contributors

• @jupyterjazz made their first contribution in #351

Full Changelog: 0.1.24...0.1.25

v0.1.24

🎉 Introduction to new functions of GPTCache

1. Support the langchain embedding

from gptcache.embedding import LangChain
from langchain.embeddings.openai import OpenAIEmbeddings

test_sentence = 'Hello, world.'
embeddings = OpenAIEmbeddings(model="your-embeddings-deployment-name")
encoder = LangChain(embeddings=embeddings)
embed = encoder.to_embeddings(test_sentence)

2. Add gptcache client

from gptcache import Client

client = Client()
client.put("Hi", "Hi back")
ans = client.get("Hi")

3. Support pgvector as vector store

from gptcache.manager import manager_factory

data_manager = manager_factory("sqlite,pgvector", vector_params={"dimension": 10})

4. Add the GPTCache server doc

reference: https://github.com/zilliztech/GPTCache/blob/main/docs/usage.md#Build-GPTCache-server

What's Changed

• Update the version to 0.1.24 by @SimFG in #347

Full Changelog: 0.1.23...0.1.24

v0.1.23

🎉 Introduction to new functions of GPTCache

1. Support the session for the LangChainLLMs

from langchain import OpenAI
from gptcache.adapter.langchain_models import LangChainLLMs
from gptcache.session import Session

session = Session(name="sqlite-example")
llm = LangChainLLMs(llm=OpenAI(temperature=0), session=session)

2. Optimize the summarization context process

from gptcache import cache
from gptcache.processor.context.summarization_context import SummarizationContextProcess

context_process = SummarizationContextProcess()
cache.init(
    pre_embedding_func=context_process.pre_process,
)

3. Add BabyAGI bootcamp

details: https://github.com/zilliztech/GPTCache/blob/main/docs/bootcamp/langchain/baby_agi.ipynb

What's Changed

• Update langchain llms with session by @shiyu22 in #327
• Wrap gptcache server in a docker image by @Chiiizzzy in #329
• Fix requirements conflict for sphinx by @jaelgu in #330
• Use self-hosted tokenizer and update summarization context. by @wxywb in #331
• Optimize some code by @SimFG in #333
• Add BabyAGI bootcamp by @shiyu22 in #334
• Improve the api for the import_ruamel by @SimFG in #336

Full Changelog: 0.1.22...0.1.23

v0.1.22

🎉 Introduction to new functions of GPTCache

1. Process the dialog context through the context processing interface, which currently supports two ways: summarize and selective context

import transformers
from gptcache.processor.context.summarization_context import SummarizationContextProcess
from gptcache.processor.context.selective_context import SelectiveContextProcess
from gptcache import cache

summarizer = transformers.pipeline("summarization", model="facebook/bart-large-cnn")
context_process = SummarizationContextProcess(summarizer, None, 512)
cache.init(
    pre_embedding_func=context_process.pre_process,
)

context_processor = SelectiveContextProcess()
cache.init(
    pre_embedding_func=context_process.pre_process,
)

What's Changed

• Add SummarizationContextProcess by @wxywb in #316
• Support duckdb by @SimFG in #323
• Add cmd for gptcache server by @Chiiizzzy in #325
• Support the selective context processor by @SimFG in #326

Full Changelog: 0.1.21...0.1.22

v0.1.21

🎉 Introduction to new functions of GPTCache

1. Support the temperature param

from gptcache.adapter import openai

openai.ChatCompletion.create(
    model="gpt-3.5-turbo",
    temperature = 1.0,  # Change temperature here
    messages=[{
        "role": "user",
        "content": question
    }],
)

2. Add the session layer

from gptcache.adapter import openai
from gptcache.session import Session

session = Session(name="my-session")
question = "what do you think about chatgpt"
openai.ChatCompletion.create(
    model="gpt-3.5-turbo",
    messages=[
        {"role": "user", "content": question}
    ],
    session=session
)

details: https://github.com/zilliztech/GPTCache/tree/main/examples#How-to-run-with-session

3. Support config cache with yaml for server

from gptcache.adapter.api import init_similar_cache_from_config

init_similar_cache_from_config(config_dir="cache_config_template.yml")

config file template: https://github.com/zilliztech/GPTCache/blob/main/cache_config_template.yml

4. Adapt the dolly model

from gptcache.adapter.dolly import Dolly

llm = Dolly.from_model(model="databricks/dolly-v2-3b")
llm(question)

What's Changed

• Use temperature to control possibility of skip_cache by @jaelgu in #306
• Add template for similar cache init config by @Chiiizzzy in #308
• Add dolly by @junjiejiangjjj in #311
• Add session usage doc by @shiyu22 in #310
• Add docs for temperature by @jaelgu in #312
• Dolly and llama docs by @junjiejiangjjj in #314
• Some minor polish on cache init by @Chiiizzzy in #313
• Update the version to 0.1.21 by @SimFG in #318

Full Changelog: 0.1.20...0.1.21

v0.1.20

🎉 Introduction to new functions of GPTCache

1. support the temperature param, like openai

A non-negative number of sampling temperature, defaults to 0.
A higher temperature makes the output more random.
A lower temperature means a more deterministic and confident output.

2. Add llama adapter

from gptcache.adapter.llama_cpp import Llama

llm = Llama('./models/7B/ggml-model.bin')
answer = llm(prompt=question)

Full Changelog: 0.1.19...0.1.20

v0.1.19

🎉 Introduction to new functions of GPTCache

1. Add stability sdk adapter (text -> image)

import os
import time

from gptcache import cache
from gptcache.processor.pre import get_prompt
from gptcache.adapter.stability_sdk import StabilityInference, generation
from gptcache.embedding import Onnx
from gptcache.manager.factory import manager_factory
from gptcache.similarity_evaluation.distance import SearchDistanceEvaluation

# init gptcache
onnx = Onnx()
data_manager = manager_factory('sqlite,faiss,local', 
                               data_dir='./', 
                               vector_params={'dimension': onnx.dimension},
                               object_params={'path': './images'}
                               )
cache.init(
    pre_embedding_func=get_prompt,
    embedding_func=onnx.to_embeddings,
    data_manager=data_manager,
    similarity_evaluation=SearchDistanceEvaluation()
    )

api_key = os.getenv('STABILITY_KEY', 'key-goes-here')

stability_api = StabilityInference(
    key=os.environ['STABILITY_KEY'], # API Key reference.
    verbose=False, # Print debug messages.
    engine='stable-diffusion-xl-beta-v2-2-2', # Set the engine to use for generation.
)

start = time.time()
answers = stability_api.generate(
    prompt='a cat sitting besides a dog',
    width=256,
    height=256
    )

stability reference: https://platform.stability.ai/docs/features/text-to-image

2. Add minigpt4 adapter

Notice: It cannot be used directly, it needs to cooperate with mini-GPT4 source code, refer to: Vision-CAIR/MiniGPT-4#136

What's Changed

• Unify the format of manager variable names in manager_factory method by @SimFG in #276
• Adapt stability_sdk by @jaelgu in #277
• Add minigpt4 adapter by @shiyu22 in #274
• Update docs by @jaelgu in #278
• Make np evaluation positively correlated with the similarity. by @wxywb in #280
• Add temperature_softmax in post processor by @jaelgu in #282
• Update the version to 0.1.19 by @SimFG in #283

Full Changelog: 0.1.18...0.1.19

v0.1.18

🎉 Introduction to new functions of GPTCache

1. Add vqa bootcamp

reference: https://github.com/zilliztech/GPTCache/tree/main/docs/bootcamp/replicate

2. Add two streamlit multimodal demos

reference: https://github.com/zilliztech/GPTCache/tree/main/docs/bootcamp/streamlit

3. Add vit image embedding func

from gptcache.embedding import ViT

encoder = ViT(model="google/vit-base-patch16-384")
embed = encoder.to_embeddings(image)

4. Add init_similar_cache func for the GPTCache api module

from gptcache.adapter.api import init_similar_cache

init_similar_cache("cache_data")

5. The simple GPTCache server provides similar cache

• clone the GPTCache repo, git clone https://github.com/zilliztech/GPTCache.git
• install the gptcache model, pip install gptcache
• run the GPTCache server, cd gptcache_server && python server.py

What's Changed

• Add vqa bootcamp by @shiyu22 in #263
• Update vqa bootcamp by @shiyu22 in #265
• Modify np evaluation to support vqa by @jaelgu in #264
• Update README by @shiyu22 in #266
• Add evaluation options in vqa bootcamp by @jaelgu in #267
• Add two streamlit demos. by @wxywb in #268
• feat: Implement vit image embedding by @Pouyanpi in #270
• Add init_similar_cache func for providing easier similar cache init by @SimFG in #272
• Update sql example by @shiyu22 in #273
• Update the version to 0.1.18 by @SimFG in #275

Full Changelog: 0.1.17...0.1.18

v0.1.17

🎉 Introduction to new functions of GPTCache

1. Add image embedding timm

import requests
from PIL import Image
from gptcache.embedding import Timm

url = 'https://raw.githubusercontent.com/zilliztech/GPTCache/main/docs/GPTCache.png'
image = Image.open(requests.get(url, stream=True).raw)  # Read image url as PIL.Image      
encoder = Timm(model='resnet18')
image_tensor = encoder.preprocess(image)
embed = encoder.to_embeddings(image_tensor)

2. Add Replicate adapter, vqa (visual question answering) (experimental)

from gptcache.adapter import replicate

question = "what is in the image?"

replicate.run(
    "andreasjansson/blip-2:xxx",
    input={
        "image": open(image_path, 'rb'),
        "question": question
    }
)

3. Support to flush data for preventing accidental loss of memory data

from gptcache import cache

cache.flush()

What's Changed

• Add image embedding timm by @jaelgu in #249
• Fix typo in release_note.md by @eltociear in #251
• Allow image path or file-like object as input of image embedding by @jaelgu in #252
• Fix docs by @junjiejiangjjj in #250
• Add replicate run adapter by @shiyu22 in #248
• Fix adapter with image object store by @jaelgu in #253
• Fix image mode issue by @jaelgu in #254
• Fix that 'NoneType' object has no attribute 'answers' by @SimFG in #256
• Update the bootcamp doc by @shiyu22 in #257
• Support to flush data for preventing accidental loss of memory data by @SimFG in #258
• Add evaluator for vqa by @jaelgu in #259
• Enable adapter.replicate docstring in API reference by @jaelgu in #260
• Update the version to 0.1.17 by @SimFG in #261

New Contributors

• @eltociear made their first contribution in #251

Full Changelog: 0.1.16...0.1.17

v0.1.16

🎉 Introduction to new functions of GPTCache

1. Add StableDiffusion adapter (experimental)

import torch

from gptcache.adapter.diffusers import StableDiffusionPipeline
from gptcache.processor.pre import get_prompt
from gptcache import cache

cache.init(
    pre_embedding_func=get_prompt,
)
model_id = "stabilityai/stable-diffusion-2-1"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)

prompt = "a photo of an astronaut riding a horse on mars"
pipe(prompt=prompt).images[0]

2. Add speech to text bootcamp, link

3. More convenient management of cache files

from gptcache.manager.factory import manager_factory

data_manager = manager_factory('sqlite,faiss', data_dir="test_cache", vector_params={"dimension": 5})

4. Add a simple GPTCache server (experimental)

After starting this server, you can:

• put the data to cache, like: curl -X PUT -d "receive a hello message" "http://localhost:8000?prompt=hello"
• get the data from cache, like: curl -X GET "http://localhost:8000?prompt=hello"

Currently the service is just a map cache, more functions are still under development.

What's Changed

• Adapt StableDiffusion by @jaelgu in #234
• Add audio embedding with data2vec by @jaelgu in #238
• Support multi-model question by @junjiejiangjjj in #235
• Add speech to text bootcamp by @shiyu22 in #239
• Fix auto API references script for gptcache.adapter by @jaelgu in #240
• Update README with multimodal adapter in modules by @jaelgu in #242
• Add manager_factory to create data_manager by @junjiejiangjjj in #241
• Add a simple GPTCache server by @SimFG in #244

Full Changelog: 0.1.15...0.1.16