0.25.3 (#113)

* add gpt-4o

add gpt-4o-2024-05-13

* fix koboldcpp client jiggle arguments

* kcpp api url default port 5001

* fix repetition breaking issues with kcpp client

* use tokencount endpoint if available

* auto configure visual agent with koboldcpp

* env var config for frontend serve

* its not clear that gpt-4o is better than turbo, dont default to it yet

* 0.25.3

* handle kcpp being down during a1111 setup check

* only check a1111 setup if client is connected

* fix kcpp a1111 setup check

* fix issue where saving a new scene could cause recent config changes to revert

README.md

@@ -57,6 +57,7 @@ Please read the documents in the `docs` folder for more advanced configuration a
     - [Ready to go](#ready-to-go)
     - [Load the introductory scenario "Infinity Quest"](#load-the-introductory-scenario-infinity-quest)
     - [Loading character cards](#loading-character-cards)
+- [Configure for hosting](#configure-for-hosting)
 - [Text-to-Speech (TTS)](docs/tts.md)
 - [Visual Generation](docs/visual.md)
 - [ChromaDB (long term memory) configuration](docs/chromadb.md)
@@ -252,3 +253,17 @@ Expand the "Load" menu in the top left corner and either click on "Upload a char
 Once a character is uploaded, talemate may actually take a moment because it needs to convert it to a talemate format and will also run additional LLM prompts to generate character attributes and world state.
 
 Make sure you save the scene after the character is loaded as it can then be loaded as normal talemate scenario in the future.
+
+## Configure for hosting
+
+By default talemate is configured to run locally. If you want to host it behind a reverse proxy or on a server, you will need create some environment variables in the `talemate_frontend/.env.development.local` file
+
+Start by copying `talemate_frontend/example.env.development.local` to `talemate_frontend/.env.development.local`.
+
+Then open the file and edit the `ALLOWED_HOSTS` and  `VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL` variables.
+
+```sh
+ALLOWED_HOSTS=example.com
+# wss if behind ssl, ws if not
+VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL=wss://example.com:5050
+```

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "talemate"
-version = "0.25.2"
+version = "0.25.3"
 description = "AI-backed roleplay and narrative tools"
 authors = ["FinalWombat"]
 license = "GNU Affero General Public License v3.0"

src/talemate/__init__.py

@@ -2,4 +2,4 @@ from .agents import Agent
 from .client import TextGeneratorWebuiClient
 from .tale_mate import *
 
-VERSION = "0.25.2"
+VERSION = "0.25.3"

src/talemate/agents/base.py

@@ -221,6 +221,9 @@ class Agent(ABC):
         if callback:
             await callback()
 
+    async def setup_check(self):
+        return False
+
     async def ready_check(self, task: asyncio.Task = None):
         self.ready_check_error = None
         if task:

src/talemate/agents/visual/__init__.py

@@ -80,6 +80,11 @@ class VisualBase(Agent):
                     ),
                 },
             ),
+            "automatic_setup": AgentAction(
+                enabled=True,
+                label="Automatic Setup",
+                description="Automatically setup the visual agent if the selected client has an implementation of the selected backend. (Like the KoboldCpp Automatic1111 api)",
+            ),
             "automatic_generation": AgentAction(
                 enabled=False,
                 label="Automatic Generation",
@@ -187,8 +192,10 @@ class VisualBase(Agent):
         prev_ready = self.backend_ready
         self.backend_ready = False
         self.ready_check_error = str(error)
+        await self.setup_check()
         if prev_ready:
             await self.emit_status()
+        
 
     async def ready_check(self):
         if not self.enabled:
@@ -198,6 +205,15 @@ class VisualBase(Agent):
         task = asyncio.create_task(fn())
         await super().ready_check(task)
 
+    async def setup_check(self):
+        
+        if not self.actions["automatic_setup"].enabled:
+            return
+        
+        backend = self.backend
+        if self.client and hasattr(self.client, f"visual_{backend.lower()}_setup"):
+            await getattr(self.client, f"visual_{backend.lower()}_setup")(self)
+
     async def apply_config(self, *args, **kwargs):
 
         try:

src/talemate/client/base.py

@@ -122,6 +122,10 @@ class ClientBase:
         """
         return self.Meta().requires_prompt_template
 
+    @property
+    def max_tokens_param_name(self):
+        return "max_tokens"
+
     def set_client(self, **kwargs):
         self.client = AsyncOpenAI(base_url=self.api_url, api_key="sk-1111")
 
@@ -625,7 +629,7 @@ class ClientBase:
             is_repetition, similarity_score, matched_line = util.similarity_score(
                 response, finalized_prompt.split("\n"), similarity_threshold=80
             )
-
+            
             if not is_repetition:
                 # not a repetition, return the response
 
@@ -659,7 +663,7 @@ class ClientBase:
 
                 # then we pad the max_tokens by the pad_max_tokens amount
 
-                prompt_param["max_tokens"] += pad_max_tokens
+                prompt_param[self.max_tokens_param_name] += pad_max_tokens
 
                 # send the prompt again
                 # we use the repetition_adjustment method to further encourage
@@ -681,7 +685,7 @@ class ClientBase:
 
                 # a lot of the times the response will now contain the repetition + something new
                 # so we dedupe the response to remove the repetition on sentences level
-
+                
                 response = util.dedupe_sentences(
                     response, matched_line, similarity_threshold=85, debug=True
                 )

src/talemate/client/koboldccp.py

@@ -1,18 +1,24 @@
 import random
 import re
+from typing import TYPE_CHECKING
 
 # import urljoin
-from urllib.parse import urljoin
+from urllib.parse import urljoin, urlparse
 import httpx
 import structlog
 
 from talemate.client.base import STOPPING_STRINGS, ClientBase, Defaults, ExtraField
 from talemate.client.registry import register
+import talemate.util as util
+
+if TYPE_CHECKING:
+    from talemate.agents.visual import VisualBase
 
 log = structlog.get_logger("talemate.client.koboldcpp")
 
 
 class KoboldCppClientDefaults(Defaults):
+    api_url: str = "http://localhost:5001"
     api_key: str = ""
 
 
@@ -35,6 +41,11 @@ class KoboldCppClient(ClientBase):
             headers["Authorization"] = f"Bearer {self.api_key}"
         return headers
 
+    @property
+    def url(self) -> str:
+        parts = urlparse(self.api_url)
+        return f"{parts.scheme}://{parts.netloc}"
+
     @property
     def is_openai(self) -> bool:
         """
@@ -63,13 +74,20 @@ class KoboldCppClient(ClientBase):
             # join /api/v1/generate
             return urljoin(self.api_url, "generate")
 
+    @property
+    def max_tokens_param_name(self):
+        if self.is_openai:
+            return "max_tokens"
+        else:
+            return "max_length"
+
     def api_endpoint_specified(self, url: str) -> bool:
         return "/v1" in self.api_url
 
     def ensure_api_endpoint_specified(self):
         if not self.api_endpoint_specified(self.api_url):
             # url doesn't specify the api endpoint
-            # use the koboldcpp openai api
+            # use the koboldcpp united api
             self.api_url = urljoin(self.api_url.rstrip("/") + "/", "/api/v1/")
         if not self.api_url.endswith("/"):
             self.api_url += "/"
@@ -126,6 +144,9 @@ class KoboldCppClient(ClientBase):
     def set_client(self, **kwargs):
         self.api_key = kwargs.get("api_key", self.api_key)
         self.ensure_api_endpoint_specified()
+        
+        
+        
 
     async def get_model_name(self):
         self.ensure_api_endpoint_specified()
@@ -153,6 +174,35 @@ class KoboldCppClient(ClientBase):
 
         return model_name
 
+    async def tokencount(self, content:str) -> int:
+        """
+        KoboldCpp has a tokencount endpoint we can use to count tokens
+        for the prompt and response
+        
+        If the endpoint is not available, we will use the default token count estimate
+        """
+        
+        # extract scheme and host from api url
+        
+        parts = urlparse(self.api_url)
+        
+        url_tokencount = f"{parts.scheme}://{parts.netloc}/api/extra/tokencount"
+        
+        async with httpx.AsyncClient() as client:
+            response = await client.post(
+                url_tokencount,
+                json={"prompt":content},
+                timeout=None,
+                headers=self.request_headers,
+            )
+            
+            if response.status_code == 404:
+                # kobold united doesn't have tokencount endpoint
+                return util.count_tokens(content)
+            
+            tokencount = len(response.json().get("ids",[]))
+            return tokencount
+        
     async def generate(self, prompt: str, parameters: dict, kind: str):
         """
         Generates text from the given prompt and parameters.
@@ -160,6 +210,8 @@ class KoboldCppClient(ClientBase):
 
         parameters["prompt"] = prompt.strip(" ")
         
+        self._returned_prompt_tokens = await self.tokencount(parameters["prompt"] )
+        
         async with httpx.AsyncClient() as client:
             response = await client.post(
                 self.api_url_for_generation,
@@ -168,15 +220,18 @@ class KoboldCppClient(ClientBase):
                 headers=self.request_headers,
             )
             response_data = response.json()
-
             try:
                 if self.is_openai:
-                    return response_data["choices"][0]["text"]
+                    response_text = response_data["choices"][0]["text"]
                 else:
-                    return response_data["results"][0]["text"]
+                    response_text = response_data["results"][0]["text"]
             except (TypeError, KeyError) as exc:
                 log.error("Failed to generate text", exc=exc, response_data=response_data, response_status=response.status_code)
-                return ""
+                response_text = ""
+                
+            self._returned_response_tokens = await self.tokencount(response_text)
+            return response_text
+            
 
     def jiggle_randomness(self, prompt_config: dict, offset: float = 0.3) -> dict:
         """
@@ -185,17 +240,64 @@ class KoboldCppClient(ClientBase):
         """
 
         temp = prompt_config["temperature"]
-        rep_pen = prompt_config["rep_pen"]
+        
+        if "rep_pen" in prompt_config:
+            rep_pen_key = "rep_pen"
+        elif "frequency_penalty" in prompt_config:
+            rep_pen_key = "frequency_penalty"
+        else:
+            rep_pen_key = "repetition_penalty"
+        
+        rep_pen = prompt_config[rep_pen_key]
 
         min_offset = offset * 0.3
 
         prompt_config["temperature"] = random.uniform(temp + min_offset, temp + offset)
-        prompt_config["rep_pen"] = random.uniform(
+        prompt_config[rep_pen_key] = random.uniform(
             rep_pen + min_offset * 0.3, rep_pen + offset * 0.3
         )
-
+        
     def reconfigure(self, **kwargs):
         if "api_key" in kwargs:
             self.api_key = kwargs.pop("api_key")
 
         super().reconfigure(**kwargs)
+
+
+    async def visual_automatic1111_setup(self, visual_agent:"VisualBase") -> bool:
+        
+        """
+        Automatically configure the visual agent for automatic1111
+        if the koboldcpp server has a SD model available
+        """
+        
+        if not self.connected:
+            return False
+        
+        sd_models_url = urljoin(self.url, "/sdapi/v1/sd-models")
+        
+        async with httpx.AsyncClient() as client:
+            
+            try:
+                response = await client.get(
+                   url=sd_models_url, timeout=2
+                )
+            except Exception as exc:
+                log.error(f"Failed to fetch sd models from {sd_models_url}", exc=exc)
+                return False
+            
+            if response.status_code != 200:
+                return False
+            
+            response_data = response.json()
+            
+            sd_model = response_data[0].get("model_name") if response_data else None
+            
+        log.info("automatic1111_setup", sd_model=sd_model)
+        if not sd_model:
+            return False
+        
+        visual_agent.actions["automatic1111"].config["api_url"].value = self.url
+        visual_agent.is_enabled = True
+        return True
+        

src/talemate/client/openai.py

@@ -28,12 +28,14 @@ SUPPORTED_MODELS = [
     "gpt-4-turbo-preview",
     "gpt-4-turbo-2024-04-09",
     "gpt-4-turbo",
+    "gpt-4o-2024-05-13",
+    "gpt-4o",
 ]
 
+# any model starting with gpt-4- is assumed to support 'json_object'
+# for others we need to explicitly state the model name
 JSON_OBJECT_RESPONSE_MODELS = [
-    "gpt-4-1106-preview",
-    "gpt-4-0125-preview",
-    "gpt-4-turbo-preview",
+    "gpt-4o",
     "gpt-3.5-turbo-0125",
 ]
 

src/talemate/instance.py

@@ -187,3 +187,5 @@ async def agent_ready_checks():
     for agent in AGENTS.values():
         if agent and agent.enabled:
             await agent.ready_check()
+        elif agent and not agent.enabled:
+            await agent.setup_check()

src/talemate/tale_mate.py

@@ -2123,7 +2123,7 @@ class Scene(Emitter):
 
     async def add_to_recent_scenes(self):
         log.debug("add_to_recent_scenes", filename=self.filename)
-        config = Config(**self.config)
+        config = load_config(as_model=True)
         config.recent_scenes.push(self)
         config.save()
 

talemate_frontend/example.env.development.local

@@ -0,0 +1,3 @@
+ALLOWED_HOSTS=example.com
+# wss if behind ssl, ws if not
+VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL=wss://example.com:5050

talemate_frontend/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "talemate_frontend",
-  "version": "0.25.2",
+  "version": "0.25.3",
   "lockfileVersion": 2,
   "requires": true,
   "packages": {
     "": {
       "name": "talemate_frontend",
-      "version": "0.25.2",
+      "version": "0.25.3",
       "dependencies": {
         "@codemirror/lang-markdown": "^6.2.5",
         "@codemirror/theme-one-dark": "^6.1.2",

talemate_frontend/package.json

@@ -1,6 +1,6 @@
 {
   "name": "talemate_frontend",
-  "version": "0.25.2",
+  "version": "0.25.3",
   "private": true,
   "scripts": {
     "serve": "vue-cli-service serve",

talemate_frontend/src/components/TalemateApp.vue

@@ -303,9 +303,11 @@ export default {
 
       this.connecting = true;
       let currentUrl = new URL(window.location.href);
-      console.log(currentUrl);
+      let websocketUrl = process.env.VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL || `ws://${currentUrl.hostname}:5050/ws`;
 
-      this.websocket = new WebSocket(`ws://${currentUrl.hostname}:5050/ws`);
+      console.log("urls", { websocketUrl, currentUrl }, {env : process.env});
+
+      this.websocket = new WebSocket(websocketUrl);
       console.log("Websocket connecting ...")
       this.websocket.onmessage = this.handleMessage;
       this.websocket.onopen = () => {

talemate_frontend/vue.config.js

@@ -1,4 +1,16 @@
 const { defineConfig } = require('@vue/cli-service')
+
+const ALLOWED_HOSTS = process.env.ALLOWED_HOSTS || "all"
+const VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL = process.env.VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL || null
+
+// if ALLOWED_HOSTS is set and has , then split it
+if (ALLOWED_HOSTS !== "all") {
+  ALLOWED_HOSTS = ALLOWED_HOSTS.split(",")
+}
+
+console.log("ALLOWED_HOSTS", ALLOWED_HOSTS)
+console.log("VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL", VUE_APP_TALEMATE_BACKEND_WEBSOCKET_URL)
+
 module.exports = defineConfig({
   transpileDependencies: true,
 
@@ -9,7 +21,7 @@ module.exports = defineConfig({
   },
 
   devServer: {
-    allowedHosts: "all",
+    allowedHosts: ALLOWED_HOSTS,
     client: {
       overlay: {
         warnings: false,