feat: add bert.cpp embeddings (#222)

Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com>

.github/workflows/bump_deps.yaml

@@ -21,6 +21,12 @@ jobs:
           - repository: "donomii/go-rwkv.cpp"
             variable: "RWKV_VERSION"
             branch: "main"
+          - repository: "ggerganov/whisper.cpp"
+            variable: "WHISPER_CPP_VERSION"
+            branch: "master"
+          - repository: "go-skynet/go-bert.cpp"
+            variable: "BERT_VERSION"
+            branch: "master"
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v3

.gitignore

@@ -3,6 +3,7 @@ go-llama
 go-gpt4all-j
 go-gpt2
 go-rwkv
+whisper.cpp
 
 # LocalAI build binary
 LocalAI
@@ -12,4 +13,7 @@ local-ai
 
 # Ignore models
 models/*
-test-models/
+test-models/
+
+# just in case
+.DS_Store

Makefile

@@ -3,11 +3,14 @@ GOTEST=$(GOCMD) test
 GOVET=$(GOCMD) vet
 BINARY_NAME=local-ai
 
-GOLLAMA_VERSION?=cf9b522db63898dcc5eb86e37c979ab85cbd583e
+GOLLAMA_VERSION?=c03e8adbc45c866e0f6d876af1887d6b01d57eb4
 GOGPT4ALLJ_VERSION?=1f7bff57f66cb7062e40d0ac3abd2217815e5109
 GOGPT2_VERSION?=245a5bfe6708ab80dc5c733dcdbfbe3cfd2acdaa
 RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
 RWKV_VERSION?=af62fcc432be2847acb6e0688b2c2491d6588d58
+WHISPER_CPP_VERSION?=bf2449dfae35a46b2cd92ab22661ce81a48d4993
+BERT_VERSION?=ec771ec715576ac050263bb7bb74bfd616a5ba13
+
 
 GREEN  := $(shell tput -Txterm setaf 2)
 YELLOW := $(shell tput -Txterm setaf 3)
@@ -15,8 +18,8 @@ WHITE  := $(shell tput -Txterm setaf 7)
 CYAN   := $(shell tput -Txterm setaf 6)
 RESET  := $(shell tput -Txterm sgr0)
 
-C_INCLUDE_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv
-LIBRARY_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv
+C_INCLUDE_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert
+LIBRARY_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert
 
 # Use this if you want to set the default behavior
 ifndef BUILD_TYPE
@@ -47,6 +50,14 @@ go-gpt4all-j:
 	@find ./go-gpt4all-j -type f -name "*.cpp" -exec sed -i'' -e 's/void replace/void json_gptj_replace/g' {} +
 	@find ./go-gpt4all-j -type f -name "*.cpp" -exec sed -i'' -e 's/::replace/::json_gptj_replace/g' {} +
 
+## BERT embeddings
+go-bert:
+	git clone --recurse-submodules https://github.com/go-skynet/go-bert.cpp go-bert
+	cd go-bert && git checkout -b build $(BERT_VERSION) && git submodule update --init --recursive --depth 1
+	@find ./go-bert -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_bert_/g' {} +
+	@find ./go-bert -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_bert_/g' {} +
+	@find ./go-bert -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_bert_/g' {} +
+
 ## RWKV
 go-rwkv:
 	git clone --recurse-submodules $(RWKV_REPO) go-rwkv
@@ -58,6 +69,9 @@ go-rwkv:
 go-rwkv/librwkv.a: go-rwkv
 	cd go-rwkv && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a .. && cp ggml/src/libggml.a ..
 
+go-bert/libgobert.a: go-bert
+	$(MAKE) -C go-bert libgobert.a
+
 go-gpt4all-j/libgptj.a: go-gpt4all-j
 	$(MAKE) -C go-gpt4all-j $(GENERIC_PREFIX)libgptj.a
 
@@ -76,6 +90,13 @@ go-gpt2:
 go-gpt2/libgpt2.a: go-gpt2
 	$(MAKE) -C go-gpt2 $(GENERIC_PREFIX)libgpt2.a
 
+whisper.cpp:
+	git clone https://github.com/ggerganov/whisper.cpp.git
+	cd whisper.cpp && git checkout -b build $(WHISPER_CPP_VERSION) && git submodule update --init --recursive --depth 1
+
+whisper.cpp/libwhisper.a: whisper.cpp
+	cd whisper.cpp && make libwhisper.a
+
 go-llama:
 	git clone --recurse-submodules https://github.com/go-skynet/go-llama.cpp go-llama
 	cd go-llama && git checkout -b build $(GOLLAMA_VERSION) && git submodule update --init --recursive --depth 1
@@ -88,8 +109,10 @@ replace:
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-gpt4all-j.cpp=$(shell pwd)/go-gpt4all-j
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-gpt2.cpp=$(shell pwd)/go-gpt2
 	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(shell pwd)/go-rwkv
+	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(shell pwd)/whisper.cpp
+	$(GOCMD) mod edit -replace github.com/go-skynet/go-bert.cpp=$(shell pwd)/go-bert
 
-prepare-sources: go-llama go-gpt2 go-gpt4all-j go-rwkv
+prepare-sources: go-llama go-gpt2 go-gpt4all-j go-rwkv whisper.cpp go-bert
 	$(GOCMD) mod download
 
 ## GENERIC
@@ -98,15 +121,18 @@ rebuild: ## Rebuilds the project
 	$(MAKE) -C go-gpt4all-j clean
 	$(MAKE) -C go-gpt2 clean
 	$(MAKE) -C go-rwkv clean
+	$(MAKE) -C whisper.cpp clean
+	$(MAKE) -C go-bert clean
 	$(MAKE) build
 
-prepare: prepare-sources go-llama/libbinding.a go-gpt4all-j/libgptj.a go-gpt2/libgpt2.a go-rwkv/librwkv.a replace ## Prepares for building
+prepare: prepare-sources go-llama/libbinding.a go-gpt4all-j/libgptj.a go-bert/libgobert.a go-gpt2/libgpt2.a go-rwkv/librwkv.a whisper.cpp/libwhisper.a replace ## Prepares for building
 
 clean: ## Remove build related file
 	rm -fr ./go-llama
 	rm -rf ./go-gpt4all-j
 	rm -rf ./go-gpt2
 	rm -rf ./go-rwkv
+	rm -rf ./go-bert
 	rm -rf $(BINARY_NAME)
 
 ## Build:
@@ -130,7 +156,7 @@ test-models/testmodel:
 
 test: prepare test-models/testmodel
 	cp tests/fixtures/* test-models
-	@C_INCLUDE_PATH=${C_INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models $(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo -v -r ./...
+	@C_INCLUDE_PATH=${C_INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models $(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo -v -r ./api
 
 ## Help:
 help: ## Show this help.

README.md

@@ -9,7 +9,7 @@
 
 [![](https://dcbadge.vercel.app/api/server/uJAeKSAGDy?style=flat-square&theme=default-inverted)](https://discord.gg/uJAeKSAGDy) 
 
-**LocalAI** is a drop-in replacement REST API compatible with OpenAI for local CPU inferencing. It allows to run models locally or on-prem with consumer grade hardware. It is based on [llama.cpp](https://github.com/ggerganov/llama.cpp), [gpt4all](https://github.com/nomic-ai/gpt4all), [rwkv.cpp](https://github.com/saharNooby/rwkv.cpp) and [ggml](https://github.com/ggerganov/ggml), including support GPT4ALL-J which is licensed under Apache 2.0.
+**LocalAI** is a drop-in replacement REST API compatible with OpenAI for local CPU inferencing. It allows to run models locally or on-prem with consumer grade hardware, supporting multiple models families. Supports also GPT4ALL-J which is licensed under Apache 2.0.
 
 - OpenAI compatible API
 - Supports multiple models
@@ -19,10 +19,15 @@
 
 LocalAI is a community-driven project, focused on making the AI accessible to anyone. Any contribution, feedback and PR is welcome! It was initially created by [mudler](https://github.com/mudler/) at the [SpectroCloud OSS Office](https://github.com/spectrocloud).
 
+LocalAI uses C++ bindings for optimizing speed. It is based on [llama.cpp](https://github.com/ggerganov/llama.cpp), [gpt4all](https://github.com/nomic-ai/gpt4all), [rwkv.cpp](https://github.com/saharNooby/rwkv.cpp), [ggml](https://github.com/ggerganov/ggml), [whisper.cpp](https://github.com/ggerganov/whisper.cpp) for audio transcriptions, and [bert.cpp](https://github.com/skeskinen/bert.cpp) for embedding.
+
 See [examples on how to integrate LocalAI](https://github.com/go-skynet/LocalAI/tree/master/examples/).
 
 ## News
 
+- 10-05-2023: Added support for fast and accurate embeddings with `bert.cpp` ( https://github.com/go-skynet/LocalAI/pull/222 )
+- 09-05-2023: Added experimental support for transcriptions endpoint ( https://github.com/go-skynet/LocalAI/pull/211 )
+- 08-05-2023: Support for embeddings with models using the `llama.cpp` backend ( https://github.com/go-skynet/LocalAI/pull/207 )
 - 02-05-2023: Support for `rwkv.cpp` models ( https://github.com/go-skynet/LocalAI/pull/158 ) and for `/edits` endpoint
 - 01-05-2023: Support for SSE stream of tokens in `llama.cpp` backends ( https://github.com/go-skynet/LocalAI/pull/152 )
 
@@ -529,6 +534,50 @@ curl http://localhost:8080/v1/models
 
 </details>
 
+### Embeddings
+
+<details>
+
+The embedding endpoint is experimental and enabled only if the model is configured with `embeddings: true` in its `yaml` file, for example:
+
+```yaml
+name: text-embedding-ada-002
+parameters:
+  model: bert
+embeddings: true
+```
+
+There is an example available [here](https://github.com/go-skynet/LocalAI/tree/master/examples/query_data/).
+
+Note: embeddings is supported only with `llama.cpp` compatible models and `bert` models. bert is more performant and available independently of the LLM model.
+
+</details>
+
+### Transcriptions endpoint
+
+<details>
+
+Note: requires ffmpeg in the container image, which is currently not shipped due to licensing issues. We will prepare separated images with ffmpeg. (stay tuned!)
+
+Download one of the models from https://huggingface.co/ggerganov/whisper.cpp/tree/main in the `models` folder, and create a YAML file for your model:
+
+```yaml
+name: whisper-1
+parameters:
+  model: whisper-en
+```
+
+The transcriptions endpoint then can be tested like so:
+```
+wget --quiet --show-progress -O gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
+
+curl http://localhost:8080/v1/audio/transcriptions -H "Content-Type: multipart/form-data" -F file="@$PWD/gb1.ogg" -F model="whisper-1"                                                     
+
+{"text":"My fellow Americans, this day has brought terrible news and great sadness to our country.At nine o'clock this morning, Mission Control in Houston lost contact with our Space ShuttleColumbia.A short time later, debris was seen falling from the skies above Texas.The Columbia's lost.There are no survivors.One board was a crew of seven.Colonel Rick Husband, Lieutenant Colonel Michael Anderson, Commander Laurel Clark, Captain DavidBrown, Commander William McCool, Dr. Kultna Shavla, and Elon Ramon, a colonel in the IsraeliAir Force.These men and women assumed great risk in the service to all humanity.In an age when spaceflight has come to seem almost routine, it is easy to overlook thedangers of travel by rocket and the difficulties of navigating the fierce outer atmosphere ofthe Earth.These astronauts knew the dangers, and they faced them willingly, knowing they had a highand noble purpose in life.Because of their courage and daring and idealism, we will miss them all the more.All Americans today are thinking as well of the families of these men and women who havebeen given this sudden shock and grief.You're not alone.Our entire nation agrees with you, and those you loved will always have the respect andgratitude of this country.The cause in which they died will continue.Mankind has led into the darkness beyond our world by the inspiration of discovery andthe longing to understand.Our journey into space will go on.In the skies today, we saw destruction and tragedy.As farther than we can see, there is comfort and hope.In the words of the prophet Isaiah, \"Lift your eyes and look to the heavens who createdall these, he who brings out the starry hosts one by one and calls them each by name.\"Because of his great power and mighty strength, not one of them is missing.The same creator who names the stars also knows the names of the seven souls we mourntoday.The crew of the shuttle Columbia did not return safely to Earth yet we can pray that all aresafely home.May God bless the grieving families and may God continue to bless America.[BLANK_AUDIO]"}
+```
+
+</details>
+
 ## Frequently asked questions
 
 Here are answers to some of the most common questions.
@@ -590,6 +639,7 @@ Feel free to open up a PR to get your project listed!
 
 - [Kairos](https://github.com/kairos-io/kairos)
 - [k8sgpt](https://github.com/k8sgpt-ai/k8sgpt#running-local-models)
+- [Spark](https://github.com/cedriking/spark)
 
 ## Blog posts and other articles
 
@@ -621,6 +671,7 @@ MIT
 - [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - [go-skynet/go-gpt4all-j.cpp](https://github.com/go-skynet/go-gpt4all-j.cpp)
 - [go-skynet/go-gpt2.cpp](https://github.com/go-skynet/go-gpt2.cpp)
+- [go-skynet/go-bert.cpp](https://github.com/go-skynet/go-bert.cpp)
 - [donomii/go-rwkv.cpp](https://github.com/donomii/go-rwkv.cpp)
 
 ## Acknowledgements

api/api.go

@@ -84,6 +84,8 @@ func App(configFile string, loader *model.ModelLoader, threads, ctxSize int, f16
 
 	app.Post("/v1/engines/:model/embeddings", embeddingsEndpoint(cm, debug, loader, threads, ctxSize, f16))
 
+	app.Post("/v1/audio/transcriptions", transcriptEndpoint(cm, debug, loader, threads, ctxSize, f16))
+
 	app.Get("/v1/models", listModels(loader, cm))
 	app.Get("/models", listModels(loader, cm))
 

api/api_test.go

@@ -79,7 +79,7 @@ var _ = Describe("API test", func() {
 		It("returns errors", func() {
 			_, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "foomodel", Prompt: "abcdedfghikl"})
 			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("error, status code: 500, message: could not load model - all backends returned error: 5 errors occurred:"))
+			Expect(err.Error()).To(ContainSubstring("error, status code: 500, message: could not load model - all backends returned error: 6 errors occurred:"))
 		})
 
 	})

api/config.go

@@ -33,6 +33,7 @@ type Config struct {
 	Mirostat       int               `yaml:"mirostat"`
 
 	PromptStrings, InputStrings []string
+	InputToken                  [][]int
 }
 
 type TemplateConfig struct {
@@ -186,8 +187,15 @@ func updateConfig(config *Config, input *OpenAIRequest) {
 		}
 	case []interface{}:
 		for _, pp := range inputs {
-			if s, ok := pp.(string); ok {
-				config.InputStrings = append(config.InputStrings, s)
+			switch i := pp.(type) {
+			case string:
+				config.InputStrings = append(config.InputStrings, i)
+			case []interface{}:
+				tokens := []int{}
+				for _, ii := range i {
+					tokens = append(tokens, int(ii.(float64)))
+				}
+				config.InputToken = append(config.InputToken, tokens)
 			}
 		}
 	}

api/openai.go

@@ -5,10 +5,16 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"net/http"
+	"os"
+	"path"
+	"path/filepath"
 	"strings"
 
 	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/whisper"
 	"github.com/gofiber/fiber/v2"
+	"github.com/otiai10/copy"
 	"github.com/rs/zerolog/log"
 	"github.com/valyala/fasthttp"
 )
@@ -69,6 +75,11 @@ type OpenAIModel struct {
 type OpenAIRequest struct {
 	Model string `json:"model" yaml:"model"`
 
+	// whisper
+	File           string `json:"file" validate:"required"`
+	ResponseFormat string `json:"response_format"`
+	Language       string `json:"language"`
+
 	// Prompt is read only by completion API calls
 	Prompt interface{} `json:"prompt" yaml:"prompt"`
 
@@ -177,10 +188,23 @@ func embeddingsEndpoint(cm ConfigMerger, debug bool, loader *model.ModelLoader,
 		log.Debug().Msgf("Parameter Config: %+v", config)
 		items := []Item{}
 
-		for i, s := range config.InputStrings {
-
+		for i, s := range config.InputToken {
 			// get the model function to call for the result
-			embedFn, err := ModelEmbedding(s, loader, *config)
+			embedFn, err := ModelEmbedding("", s, loader, *config)
+			if err != nil {
+				return err
+			}
+
+			embeddings, err := embedFn()
+			if err != nil {
+				return err
+			}
+			items = append(items, Item{Embedding: embeddings, Index: i, Object: "embedding"})
+		}
+
+		for i, s := range config.InputStrings {
+			// get the model function to call for the result
+			embedFn, err := ModelEmbedding(s, []int{}, loader, *config)
 			if err != nil {
 				return err
 			}
@@ -372,6 +396,45 @@ func editEndpoint(cm ConfigMerger, debug bool, loader *model.ModelLoader, thread
 	}
 }
 
+// https://platform.openai.com/docs/api-reference/audio/create
+func transcriptEndpoint(cm ConfigMerger, debug bool, loader *model.ModelLoader, threads, ctx int, f16 bool) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		config, input, err := readConfig(cm, c, loader, debug, threads, ctx, f16)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		// retrieve the file data from the request
+		file, err := c.FormFile("file")
+		if err != nil {
+			return c.Status(http.StatusBadRequest).JSON(fiber.Map{"error": err.Error()})
+		}
+		log.Debug().Msgf("Audio file: %+v", file)
+
+		dir, err := os.MkdirTemp("", "whisper")
+		if err != nil {
+			return err
+		}
+		defer os.RemoveAll(dir)
+
+		dst := filepath.Join(dir, path.Base(file.Filename))
+		if err := copy.Copy(file.Filename, dst); err != nil {
+			return err
+		}
+
+		log.Debug().Msgf("Audio file copied to: %+v", dst)
+
+		tr, err := whisper.Transcript(filepath.Join(loader.ModelPath, config.Model), dst, input.Language)
+		if err != nil {
+			return err
+		}
+
+		log.Debug().Msgf("Trascribed: %+v", tr)
+		// TODO: handle different outputs here
+		return c.Status(http.StatusOK).JSON(fiber.Map{"text": tr})
+	}
+}
+
 func listModels(loader *model.ModelLoader, cm ConfigMerger) func(ctx *fiber.Ctx) error {
 	return func(c *fiber.Ctx) error {
 		models, err := loader.ListModels()

api/prediction.go

@@ -8,6 +8,7 @@ import (
 
 	"github.com/donomii/go-rwkv.cpp"
 	model "github.com/go-skynet/LocalAI/pkg/model"
+	bert "github.com/go-skynet/go-bert.cpp"
 	gpt2 "github.com/go-skynet/go-gpt2.cpp"
 	gptj "github.com/go-skynet/go-gpt4all-j.cpp"
 	llama "github.com/go-skynet/go-llama.cpp"
@@ -32,7 +33,7 @@ func defaultLLamaOpts(c Config) []llama.ModelOption {
 	return llamaOpts
 }
 
-func ModelEmbedding(s string, loader *model.ModelLoader, c Config) (func() ([]float32, error), error) {
+func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, c Config) (func() ([]float32, error), error) {
 	if !c.Embeddings {
 		return nil, fmt.Errorf("endpoint disabled for this model by API configuration")
 	}
@@ -57,8 +58,19 @@ func ModelEmbedding(s string, loader *model.ModelLoader, c Config) (func() ([]fl
 	case *llama.LLama:
 		fn = func() ([]float32, error) {
 			predictOptions := buildLLamaPredictOptions(c)
+			if len(tokens) > 0 {
+				return model.TokenEmbeddings(tokens, predictOptions...)
+			}
 			return model.Embeddings(s, predictOptions...)
 		}
+	// bert embeddings
+	case *bert.Bert:
+		fn = func() ([]float32, error) {
+			if len(tokens) > 0 {
+				return nil, fmt.Errorf("embeddings endpoint for this model supports only string")
+			}
+			return model.Embeddings(s, bert.SetThreads(c.Threads))
+		}
 	default:
 		fn = func() ([]float32, error) {
 			return nil, fmt.Errorf("embeddings not supported by the backend")

examples/README.md

@@ -73,6 +73,14 @@ Shows how to integrate with [Llama-Index](https://gpt-index.readthedocs.io/en/st
 
 [Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/query_data/)
 
+### Template for Runpod.io
+
+_by [@fHachenberg](https://github.com/fHachenberg)_
+
+Allows to run any LocalAI-compatible model as a backend on the servers of https://runpod.io
+
+[Check it out here](https://runpod.io/gsc?template=uv9mtqnrd0&ref=984wlcra)
+
 ## Want to contribute?
 
 Create an issue, and put `Example: <description>` in the title! We will post your examples here.

examples/langchain-python/README.md

@@ -30,4 +30,6 @@ export OPENAI_API_KEY=sk-
 
 python test.py
 # A good company name for a company that makes colorful socks would be "Colorsocks".
+
+python agent.py
 ```

examples/langchain-python/agent.py

@@ -0,0 +1,44 @@
+## This is a fork/based from https://gist.github.com/wiseman/4a706428eaabf4af1002a07a114f61d6
+
+from io import StringIO
+import sys
+import os
+from typing import Dict, Optional
+
+from langchain.agents import load_tools
+from langchain.agents import initialize_agent
+from langchain.agents.tools import Tool
+from langchain.llms import OpenAI
+
+base_path = os.environ.get('OPENAI_API_BASE', 'http://localhost:8080/v1')
+model_name = os.environ.get('MODEL_NAME', 'gpt-3.5-turbo')
+
+class PythonREPL:
+    """Simulates a standalone Python REPL."""
+
+    def __init__(self):
+        pass        
+
+    def run(self, command: str) -> str:
+        """Run command and returns anything printed."""
+        old_stdout = sys.stdout
+        sys.stdout = mystdout = StringIO()
+        try:
+            exec(command, globals())
+            sys.stdout = old_stdout
+            output = mystdout.getvalue()
+        except Exception as e:
+            sys.stdout = old_stdout
+            output = str(e)
+        return output
+
+llm = OpenAI(temperature=0.0, openai_api_base=base_path, model_name=model_name)
+python_repl = Tool(
+        "Python REPL",
+        PythonREPL().run,
+        """A Python shell. Use this to execute python commands. Input should be a valid python command.
+        If you expect output it should be printed out.""",
+    )
+tools = [python_repl]
+agent = initialize_agent(tools, llm, agent="zero-shot-react-description", verbose=True)
+agent.run("What is the 10th fibonacci number?")

examples/query_data/README.md

@@ -4,13 +4,15 @@ This example makes use of [Llama-Index](https://gpt-index.readthedocs.io/en/stab
 
 It loosely follows [the quickstart](https://gpt-index.readthedocs.io/en/stable/guides/primer/usage_pattern.html).
 
+Summary of the steps:
+
+- prepare the dataset (and store it into `data`)
+- prepare a vector index database to run queries on
+- run queries
+
 ## Requirements
 
-For this in order to work, you will need a model compatible with the `llama.cpp` backend. This is will not work with gpt4all.
-
-The example uses `WizardLM`. Edit the config files in `models/` accordingly to specify the model you use (change `HERE`).
-
-You will also need a training data set. Copy that over `data`.
+You will need a training data set. Copy that over `data`.
 
 ## Setup
 
@@ -22,13 +24,16 @@ git clone https://github.com/go-skynet/LocalAI
 
 cd LocalAI/examples/query_data
 
-# Copy your models, edit config files accordingly
+wget https://huggingface.co/skeskinen/ggml/resolve/main/all-MiniLM-L6-v2/ggml-model-q4_0.bin -O models/bert
+wget https://gpt4all.io/models/ggml-gpt4all-j.bin -O models/ggml-gpt4all-j
 
 # start with docker-compose
 docker-compose up -d --build
 ```
 
-### Create a storage:
+### Create a storage
+
+In this step we will create a local vector database from our document set, so later we can ask questions on it with the LLM.
 
 ```bash
 export OPENAI_API_BASE=http://localhost:8080/v1
@@ -41,9 +46,22 @@ After it finishes, a directory "storage" will be created with the vector index d
 
 ## Query
 
+We can now query the dataset. 
+
 ```bash
 export OPENAI_API_BASE=http://localhost:8080/v1
 export OPENAI_API_KEY=sk-
 
 python query.py
+```
+
+## Update
+
+To update our vector database, run `update.py`
+
+```bash
+export OPENAI_API_BASE=http://localhost:8080/v1
+export OPENAI_API_KEY=sk-
+
+python update.py
 ```

examples/query_data/models/embeddings.yaml

@@ -1,18 +1,6 @@
 name: text-embedding-ada-002
 parameters:
-  model: HERE
-  top_k: 80
-  temperature: 0.2
-  top_p: 0.7
-context_size: 1024
+  model: bert
 threads: 14
-stopwords:
-- "HUMAN:"
-- "GPT:"
-roles:
-  user: " "
-  system: " "
+backend: bert-embeddings
 embeddings: true
-template:
-  completion: completion
-  chat: gpt4all

examples/query_data/models/gpt-3.5-turbo.yaml

@@ -1,12 +1,11 @@
 name: gpt-3.5-turbo
 parameters:
-  model: HERE
+  model: ggml-gpt4all-j
   top_k: 80
   temperature: 0.2
   top_p: 0.7
 context_size: 1024
 threads: 14
-embeddings: true
 stopwords:
 - "HUMAN:"
 - "GPT:"
@@ -15,4 +14,4 @@ roles:
   system: " "
 template:
   completion: completion
-  chat: wizardlm
+  chat: gpt4all

examples/query_data/models/wizardlm.tmpl

@@ -1,3 +0,0 @@
-{{.Input}}
-
-### Response:

examples/query_data/query.py

@@ -10,10 +10,10 @@ from llama_index import StorageContext, load_index_from_storage
 base_path = os.environ.get('OPENAI_API_BASE', 'http://localhost:8080/v1')
 
 # This example uses text-davinci-003 by default; feel free to change if desired
-llm_predictor = LLMPredictor(llm=OpenAI(temperature=0, model_name="gpt-3.5-turbo",openai_api_base=base_path))
+llm_predictor = LLMPredictor(llm=OpenAI(temperature=0, model_name="gpt-3.5-turbo", openai_api_base=base_path))
 
 # Configure prompt parameters and initialise helper
-max_input_size = 1024
+max_input_size = 500
 num_output = 256
 max_chunk_overlap = 20
 
@@ -29,5 +29,7 @@ storage_context = StorageContext.from_defaults(persist_dir='./storage')
 index = load_index_from_storage(storage_context,     service_context=service_context,    )
 
 query_engine = index.as_query_engine()
-response = query_engine.query("XXXXXX your question here XXXXX")
-print(response)
+
+data = input("Question: ")
+response = query_engine.query(data)
+print(response)

examples/query_data/store.py

@@ -13,15 +13,15 @@ base_path = os.environ.get('OPENAI_API_BASE', 'http://localhost:8080/v1')
 llm_predictor = LLMPredictor(llm=OpenAI(temperature=0, model_name="gpt-3.5-turbo", openai_api_base=base_path))
 
 # Configure prompt parameters and initialise helper
-max_input_size = 256
-num_output = 256
-max_chunk_overlap = 10
+max_input_size = 400
+num_output = 400
+max_chunk_overlap = 30
 
 prompt_helper = PromptHelper(max_input_size, num_output, max_chunk_overlap)
 
 # Load documents from the 'data' directory
 documents = SimpleDirectoryReader('data').load_data()
-service_context = ServiceContext.from_defaults(llm_predictor=llm_predictor, prompt_helper=prompt_helper, chunk_size_limit = 257)
+service_context = ServiceContext.from_defaults(llm_predictor=llm_predictor, prompt_helper=prompt_helper, chunk_size_limit = 400)
 index = GPTVectorStoreIndex.from_documents(documents, service_context=service_context)
 index.storage_context.persist(persist_dir="./storage")
 

examples/query_data/update.py

@@ -0,0 +1,32 @@
+import os
+
+# Uncomment to specify your OpenAI API key here (local testing only, not in production!), or add corresponding environment variable (recommended)
+# os.environ['OPENAI_API_KEY']= ""
+
+from llama_index import   LLMPredictor, PromptHelper, SimpleDirectoryReader, ServiceContext
+from langchain.llms.openai import OpenAI
+from llama_index import StorageContext, load_index_from_storage
+
+base_path = os.environ.get('OPENAI_API_BASE', 'http://localhost:8080/v1')
+
+# This example uses text-davinci-003 by default; feel free to change if desired
+llm_predictor = LLMPredictor(llm=OpenAI(temperature=0, model_name="gpt-3.5-turbo", openai_api_base=base_path))
+
+# Configure prompt parameters and initialise helper
+max_input_size = 512
+num_output = 256
+max_chunk_overlap = 20
+
+prompt_helper = PromptHelper(max_input_size, num_output, max_chunk_overlap)
+
+# Load documents from the 'data' directory
+service_context = ServiceContext.from_defaults(llm_predictor=llm_predictor, prompt_helper=prompt_helper)
+
+# rebuild storage context
+storage_context = StorageContext.from_defaults(persist_dir='./storage')
+
+# load index
+index = load_index_from_storage(storage_context,     service_context=service_context,    )
+documents = SimpleDirectoryReader('data').load_data()
+index.refresh(documents)
+index.storage_context.persist(persist_dir="./storage")

go.mod

@@ -4,13 +4,17 @@ go 1.19
 
 require (
 	github.com/donomii/go-rwkv.cpp v0.0.0-20230503112711-af62fcc432be
+	github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230508180809-bf2449dfae35
+	github.com/go-audio/wav v1.1.0
+	github.com/go-skynet/go-bert.cpp v0.0.0-20230510101404-7bb183b147ea
 	github.com/go-skynet/go-gpt2.cpp v0.0.0-20230422085954-245a5bfe6708
 	github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c
-	github.com/go-skynet/go-llama.cpp v0.0.0-20230505100647-691d479d3675
-	github.com/gofiber/fiber/v2 v2.44.0
+	github.com/go-skynet/go-llama.cpp v0.0.0-20230509080828-f4d26f43f1d3
+	github.com/gofiber/fiber/v2 v2.45.0
 	github.com/hashicorp/go-multierror v1.1.1
 	github.com/onsi/ginkgo/v2 v2.9.4
 	github.com/onsi/gomega v1.27.6
+	github.com/otiai10/copy v1.11.0
 	github.com/otiai10/openaigo v1.1.0
 	github.com/rs/zerolog v1.29.1
 	github.com/sashabaranov/go-openai v1.9.3
@@ -26,6 +30,8 @@ require (
 	github.com/PuerkitoBio/urlesc v0.0.0-20170810143723-de5bf2ad4578 // indirect
 	github.com/andybalholm/brotli v1.0.5 // indirect
 	github.com/cpuguy83/go-md2man/v2 v2.0.2 // indirect
+	github.com/go-audio/audio v1.0.0 // indirect
+	github.com/go-audio/riff v1.0.0 // indirect
 	github.com/go-logr/logr v1.2.4 // indirect
 	github.com/go-openapi/jsonpointer v0.19.5 // indirect
 	github.com/go-openapi/jsonreference v0.19.6 // indirect
@@ -52,7 +58,7 @@ require (
 	github.com/valyala/tcplisten v1.0.0 // indirect
 	github.com/xrash/smetrics v0.0.0-20201216005158-039620a65673 // indirect
 	golang.org/x/net v0.9.0 // indirect
-	golang.org/x/sys v0.7.0 // indirect
+	golang.org/x/sys v0.8.0 // indirect
 	golang.org/x/text v0.9.0 // indirect
 	golang.org/x/tools v0.8.0 // indirect
 	gopkg.in/yaml.v2 v2.4.0 // indirect

go.sum

@@ -16,8 +16,14 @@ github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ3
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
-github.com/donomii/go-rwkv.cpp v0.0.0-20230503112711-af62fcc432be h1:3Hic97PY6hcw/SY44RuR7kyONkxd744RFeRrqckzwNQ=
-github.com/donomii/go-rwkv.cpp v0.0.0-20230503112711-af62fcc432be/go.mod h1:gWy7FIWioqYmYxkaoFyBnaKApeZVrUkHhv9EV9pz4dM=
+github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230508180809-bf2449dfae35 h1:sMg/SgnMPS/HNUO/2kGm72vl8R9TmNIwgLFr2TNwR3g=
+github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230508180809-bf2449dfae35/go.mod h1:QIjZ9OktHFG7p+/m3sMvrAJKKdWrr1fZIK0rM6HZlyo=
+github.com/go-audio/audio v1.0.0 h1:zS9vebldgbQqktK4H0lUqWrG8P0NxCJVqcj7ZpNnwd4=
+github.com/go-audio/audio v1.0.0/go.mod h1:6uAu0+H2lHkwdGsAY+j2wHPNPpPoeg5AaEFh9FlA+Zs=
+github.com/go-audio/riff v1.0.0 h1:d8iCGbDvox9BfLagY94fBynxSPHO80LmZCaOsmKxokA=
+github.com/go-audio/riff v1.0.0/go.mod h1:l3cQwc85y79NQFCRB7TiPoNiaijp6q8Z0Uv38rVG498=
+github.com/go-audio/wav v1.1.0 h1:jQgLtbqBzY7G+BM8fXF7AHUk1uHUviWS4X39d5rsL2g=
+github.com/go-audio/wav v1.1.0/go.mod h1:mpe9qfwbScEbkd8uybLuIpTgHyrISw/OTuvjUW2iGtE=
 github.com/go-logr/logr v1.2.4 h1:g01GSCwiDw2xSZfjJ2/T9M+S6pFdcNtFYsp+Y43HYDQ=
 github.com/go-logr/logr v1.2.4/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=
 github.com/go-openapi/jsonpointer v0.19.3/go.mod h1:Pl9vOtqEWErmShwVjC8pYs9cog34VGT37dQOVbmoatg=
@@ -30,18 +36,11 @@ github.com/go-openapi/spec v0.20.4/go.mod h1:faYFR1CvsJZ0mNsmsphTMSoRrNV3TEDoAM7
 github.com/go-openapi/swag v0.19.5/go.mod h1:POnQmlKehdgb5mhVOsnJFsivZCEZ/vjK9gh66Z9tfKk=
 github.com/go-openapi/swag v0.19.15 h1:D2NRCBzS9/pEY3gP9Nl8aDqGUcPFrwG2p+CNFrLyrCM=
 github.com/go-openapi/swag v0.19.15/go.mod h1:QYRuS/SOXUCsnplDa677K7+DxSOj6IPNl/eQntq43wQ=
-github.com/go-skynet/go-gpt2.cpp v0.0.0-20230422085954-245a5bfe6708 h1:cfOi4TWvQ6JsAm9Q1A8I8j9YfNy10bmIfwOiyGyU5wQ=
-github.com/go-skynet/go-gpt2.cpp v0.0.0-20230422085954-245a5bfe6708/go.mod h1:1Wj/xbkMfwQSOrhNYK178IzqQHstZbRfhx4s8p1M5VM=
-github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c h1:48I7jpLNGiQeBmF0SFVVbREh8vlG0zN13v9LH5ctXis=
-github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c/go.mod h1:5VZ9XbcINI0XcHhkcX8GPK8TplFGAzu1Hrg4tNiMCtI=
-github.com/go-skynet/go-llama.cpp v0.0.0-20230504223241-67ff6a4db244/go.mod h1:LvSQx5QAYBAMpWkbyVFFDiM1Tzj8LP55DvmUM3hbRMY=
-github.com/go-skynet/go-llama.cpp v0.0.0-20230505100647-691d479d3675 h1:plXywr95RghidIHPHl+O/zpcNXenEeS6w/6WftFNr9E=
-github.com/go-skynet/go-llama.cpp v0.0.0-20230505100647-691d479d3675/go.mod h1:LvSQx5QAYBAMpWkbyVFFDiM1Tzj8LP55DvmUM3hbRMY=
 github.com/go-task/slim-sprig v0.0.0-20230315185526-52ccab3ef572 h1:tfuBGBXKqDEevZMzYi5KSi8KkcZtzBcTgAUUtapy0OI=
 github.com/go-task/slim-sprig v0.0.0-20230315185526-52ccab3ef572/go.mod h1:9Pwr4B2jHnOSGXyyzV8ROjYa2ojvAY6HCGYYfMoC3Ls=
 github.com/godbus/dbus/v5 v5.0.4/go.mod h1:xhWf0FNVPg57R7Z0UbKHbJfkEywrmjJnf7w5xrFpKfA=
-github.com/gofiber/fiber/v2 v2.44.0 h1:Z90bEvPcJM5GFJnu1py0E1ojoerkyew3iiNJ78MQCM8=
-github.com/gofiber/fiber/v2 v2.44.0/go.mod h1:VTMtb/au8g01iqvHyaCzftuM/xmZgKOZCtFzz6CdV9w=
+github.com/gofiber/fiber/v2 v2.45.0 h1:p4RpkJT9GAW6parBSbcNFH2ApnAuW3OzaQzbOCoDu+s=
+github.com/gofiber/fiber/v2 v2.45.0/go.mod h1:DNl0/c37WLe0g92U6lx1VMQuxGUQY5V7EIaVoEsUffc=
 github.com/golang/protobuf v1.5.3 h1:KhyjKVUg7Usr/dYsdSqoFveMYd5ko72D+zANwlG1mmg=
 github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38=
 github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
@@ -82,7 +81,9 @@ github.com/onsi/ginkgo/v2 v2.9.4 h1:xR7vG4IXt5RWx6FfIjyAtsoMAtnc3C/rFXBBd2AjZwE=
 github.com/onsi/ginkgo/v2 v2.9.4/go.mod h1:gCQYp2Q+kSoIj7ykSVb9nskRSsR6PUj4AiLywzIhbKM=
 github.com/onsi/gomega v1.27.6 h1:ENqfyGeS5AX/rlXDd/ETokDz93u0YufY1Pgxuy/PvWE=
 github.com/onsi/gomega v1.27.6/go.mod h1:PIQNjfQwkP3aQAH7lf7j87O/5FiNr+ZR8+ipb+qQlhg=
-github.com/otiai10/mint v1.4.1 h1:HOVBfKP1oXIc0wWo9hZ8JLdZtyCPWqjvmFDuVZ0yv2Y=
+github.com/otiai10/copy v1.11.0 h1:OKBD80J/mLBrwnzXqGtFCzprFSGioo30JcmR4APsNwc=
+github.com/otiai10/copy v1.11.0/go.mod h1:rSaLseMUsZFFbsFGc7wCJnnkTAvdc5L6VWxPE4308Ww=
+github.com/otiai10/mint v1.5.1 h1:XaPLeE+9vGbuyEHem1JNk3bYc7KKqyI/na0/mLd/Kks=
 github.com/otiai10/openaigo v1.1.0 h1:zRvGBqZUW5PCMgdkJNsPVTBd8tOLCMTipXE5wD2pdTg=
 github.com/otiai10/openaigo v1.1.0/go.mod h1:792bx6AWTS61weDi2EzKpHHnTF4eDMAlJ5GvAk/mgPg=
 github.com/philhofer/fwd v1.1.1/go.mod h1:gk3iGcWd9+svBvR0sR+KPcfE+RNWozjowpeBVG3ZVNU=
@@ -108,7 +109,7 @@ github.com/savsgio/gotils v0.0.0-20230208104028-c358bd845dee/go.mod h1:qwtSXrKuJ
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
 github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
-github.com/stretchr/testify v1.7.0 h1:nwc3DEeHmmLAfoZucVR881uASk0Mfjw8xYJ99tb5CcY=
+github.com/stretchr/testify v1.8.1 h1:w7B6lhMri9wdJUVmEZPGGhZzrYTPvgJArz7wNPgYKsk=
 github.com/swaggo/swag v1.16.1 h1:fTNRhKstPKxcnoKsytm4sahr8FaYzUcT7i1/3nd/fBg=
 github.com/swaggo/swag v1.16.1/go.mod h1:9/LMvHycG3NFHfR6LwvikHv5iFvmPADQ359cKikGxto=
 github.com/tinylib/msgp v1.1.6/go.mod h1:75BAfg2hauQhs3qedfdDZmWAPcFMAvJE5b9rGOMufyw=
@@ -161,8 +162,8 @@ golang.org/x/sys v0.0.0-20220722155257-8c9f86f7a55f/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20220811171246-fbc7d0a398ab/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.3.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.7.0 h1:3jlCCIQZPdOYu1h8BkNvLz8Kgwtae2cagcG/VamtZRU=
-golang.org/x/sys v0.7.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.8.0 h1:EBmGv8NaZBZTWvrbjNoL6HVt+IVy3QDQpJs7VRIw3tU=
+golang.org/x/sys v0.8.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
 golang.org/x/term v0.0.0-20210927222741-03fcf44c2211/go.mod h1:jbD1KX2456YbFQfuXm/mYQcufACuNUgVhRMnK/tPxf8=
 golang.org/x/term v0.3.0/go.mod h1:q750SLmJuPmVoN1blW3UFBPREJfb1KmY3vwxfr+nFDA=
@@ -186,8 +187,9 @@ golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1/go.mod h1:I/5z698sn9Ka8T
 google.golang.org/protobuf v1.28.0 h1:w43yiav+6bVFTBQFZX0r7ipe9JQ1QsbMgHwbBziscLw=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
-gopkg.in/check.v1 v1.0.0-20200227125254-8fa46927fb4f h1:BLraFXnmrev5lT+xlilqcH8XK9/i0At2xKjWk4p6zsU=
 gopkg.in/check.v1 v1.0.0-20200227125254-8fa46927fb4f/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20200902074654-038fdea0a05b h1:QRR6H1YWRnHb4Y/HeNFCTJLFVxaq6wH4YuVdsUOr75U=
+gopkg.in/check.v1 v1.0.0-20200902074654-038fdea0a05b/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v2 v2.4.0 h1:D8xgwECY7CYvx+Y2n4sBz93Jn9JRvxdiyyo8CTfuKaY=
 gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=

pkg/model/loader.go

@@ -14,6 +14,7 @@ import (
 	"github.com/rs/zerolog/log"
 
 	rwkv "github.com/donomii/go-rwkv.cpp"
+	bert "github.com/go-skynet/go-bert.cpp"
 	gpt2 "github.com/go-skynet/go-gpt2.cpp"
 	gptj "github.com/go-skynet/go-gpt4all-j.cpp"
 	llama "github.com/go-skynet/go-llama.cpp"
@@ -22,13 +23,15 @@ import (
 type ModelLoader struct {
 	ModelPath string
 	mu        sync.Mutex
-
+	// TODO: this needs generics
 	models            map[string]*llama.LLama
 	gptmodels         map[string]*gptj.GPTJ
 	gpt2models        map[string]*gpt2.GPT2
 	gptstablelmmodels map[string]*gpt2.StableLM
 	rwkv              map[string]*rwkv.RwkvState
-	promptsTemplates  map[string]*template.Template
+	bert              map[string]*bert.Bert
+
+	promptsTemplates map[string]*template.Template
 }
 
 func NewModelLoader(modelPath string) *ModelLoader {
@@ -39,6 +42,7 @@ func NewModelLoader(modelPath string) *ModelLoader {
 		gptstablelmmodels: make(map[string]*gpt2.StableLM),
 		models:            make(map[string]*llama.LLama),
 		rwkv:              make(map[string]*rwkv.RwkvState),
+		bert:              make(map[string]*bert.Bert),
 		promptsTemplates:  make(map[string]*template.Template),
 	}
 }
@@ -156,6 +160,38 @@ func (ml *ModelLoader) LoadStableLMModel(modelName string) (*gpt2.StableLM, erro
 	return model, err
 }
 
+func (ml *ModelLoader) LoadBERT(modelName string) (*bert.Bert, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	if !ml.ExistsInModelPath(modelName) {
+		return nil, fmt.Errorf("model does not exist")
+	}
+
+	if m, ok := ml.bert[modelName]; ok {
+		log.Debug().Msgf("Model already loaded in memory: %s", modelName)
+		return m, nil
+	}
+
+	// Load the model and keep it in memory for later use
+	modelFile := filepath.Join(ml.ModelPath, modelName)
+	log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
+
+	model, err := bert.New(modelFile)
+	if err != nil {
+		return nil, err
+	}
+
+	// If there is a prompt template, load it
+	if err := ml.loadTemplateIfExists(modelName, modelFile); err != nil {
+		return nil, err
+	}
+
+	ml.bert[modelName] = model
+	return model, err
+}
+
 func (ml *ModelLoader) LoadGPT2Model(modelName string) (*gpt2.GPT2, error) {
 	ml.mu.Lock()
 	defer ml.mu.Unlock()
@@ -299,6 +335,8 @@ func (ml *ModelLoader) BackendLoader(backendString string, modelFile string, lla
 		return ml.LoadGPT2Model(modelFile)
 	case "gptj":
 		return ml.LoadGPTJModel(modelFile)
+	case "bert-embeddings":
+		return ml.LoadBERT(modelFile)
 	case "rwkv":
 		return ml.LoadRWKV(modelFile, modelFile+tokenizerSuffix, threads)
 	default:
@@ -361,5 +399,13 @@ func (ml *ModelLoader) GreedyLoader(modelFile string, llamaOpts []llama.ModelOpt
 		err = multierror.Append(err, modelerr)
 	}
 
+	model, modelerr = ml.LoadBERT(modelFile)
+	if modelerr == nil {
+		updateModels(model)
+		return model, nil
+	} else {
+		err = multierror.Append(err, modelerr)
+	}
+
 	return nil, fmt.Errorf("could not load model - all backends returned error: %s", err.Error())
 }

pkg/whisper/whisper.go

@@ -0,0 +1,93 @@
+package whisper
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+
+	"github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	wav "github.com/go-audio/wav"
+)
+
+func sh(c string) (string, error) {
+	cmd := exec.Command("/bin/sh", "-c", c)
+	cmd.Env = os.Environ()
+	o, err := cmd.CombinedOutput()
+	return string(o), err
+}
+
+// AudioToWav converts audio to wav for transcribe. It bashes out to ffmpeg
+// TODO: use https://github.com/mccoyst/ogg?
+func audioToWav(src, dst string) error {
+	out, err := sh(fmt.Sprintf("ffmpeg -i %s -format s16le -ar 16000 -ac 1 -acodec pcm_s16le %s", src, dst))
+	if err != nil {
+		return fmt.Errorf("error: %w out: %s", err, out)
+	}
+
+	return nil
+}
+
+func Transcript(modelpath, audiopath, language string) (string, error) {
+
+	dir, err := os.MkdirTemp("", "whisper")
+	if err != nil {
+		return "", err
+	}
+	defer os.RemoveAll(dir)
+
+	convertedPath := filepath.Join(dir, "converted.wav")
+
+	if err := audioToWav(audiopath, convertedPath); err != nil {
+		return "", err
+	}
+
+	// Open samples
+	fh, err := os.Open(convertedPath)
+	if err != nil {
+		return "", err
+	}
+	defer fh.Close()
+
+	// Read samples
+	d := wav.NewDecoder(fh)
+	buf, err := d.FullPCMBuffer()
+	if err != nil {
+		return "", err
+	}
+
+	data := buf.AsFloat32Buffer().Data
+
+	// Load the model
+	model, err := whisper.New(modelpath)
+	if err != nil {
+		return "", err
+	}
+	defer model.Close()
+
+	// Process samples
+	context, err := model.NewContext()
+	if err != nil {
+		return "", err
+
+	}
+
+	if language != "" {
+		context.SetLanguage(language)
+	}
+
+	if err := context.Process(data, nil); err != nil {
+		return "", err
+	}
+
+	text := ""
+	for {
+		segment, err := context.NextSegment()
+		if err != nil {
+			break
+		}
+		text += segment.Text
+	}
+
+	return text, nil
+}