diff --git a/server/utils/agents/ephemeral.js b/server/utils/agents/ephemeral.js
index 7b05b4adc834a15a8e5ed7ebb2b1f05bd902d4fb..1ab2a988a2ff48a4ceda24089cc1d7d611e470ed 100644
--- a/server/utils/agents/ephemeral.js
+++ b/server/utils/agents/ephemeral.js
@@ -99,30 +99,69 @@ class EphemeralAgentHandler extends AgentHandler {
}
}
+ /**
+ * Attempts to find a fallback provider and model to use if the workspace
+ * does not have an explicit `agentProvider` and `agentModel` set.
+ * 1. Fallback to the workspace `chatProvider` and `chatModel` if they exist.
+ * 2. Fallback to the system `LLM_PROVIDER` and try to load the the associated default model via ENV params or a base available model.
+ * 3. Otherwise, return null - will likely throw an error the user can act on.
+ * @returns {object|null} - An object with provider and model keys.
+ */
+ #getFallbackProvider() {
+ // First, fallback to the workspace chat provider and model if they exist
+ if (this.#workspace.chatProvider && this.#workspace.chatModel) {
+ return {
+ provider: this.#workspace.chatProvider,
+ model: this.#workspace.chatModel,
+ };
+ }
+
+ // If workspace does not have chat provider and model fallback
+ // to system provider and try to load provider default model
+ const systemProvider = process.env.LLM_PROVIDER;
+ const systemModel = this.providerDefault(systemProvider);
+ if (systemProvider && systemModel) {
+ return {
+ provider: systemProvider,
+ model: systemModel,
+ };
+ }
+
+ return null;
+ }
+
/**
* Finds or assumes the model preference value to use for API calls.
* If multi-model loading is supported, we use their agent model selection of the workspace
* If not supported, we attempt to fallback to the system provider value for the LLM preference
* and if that fails - we assume a reasonable base model to exist.
- * @returns {string} the model preference value to use in API calls
+ * @returns {string|null} the model preference value to use in API calls
*/
#fetchModel() {
- if (!Object.keys(this.noProviderModelDefault).includes(this.provider))
- return this.#workspace.agentModel || this.providerDefault();
+ // Provider was not explicitly set for workspace, so we are going to run our fallback logic
+ // that will set a provider and model for us to use.
+ if (!this.provider) {
+ const fallback = this.#getFallbackProvider();
+ if (!fallback) throw new Error("No valid provider found for the agent.");
+ this.provider = fallback.provider; // re-set the provider to the fallback provider so it is not null.
+ return fallback.model; // set its defined model based on fallback logic.
+ }
- // Provider has no reliable default (cant load many models) - so we need to look at system
- // for the model param.
- const sysModelKey = this.noProviderModelDefault[this.provider];
- if (!!sysModelKey)
- return process.env[sysModelKey] ?? this.providerDefault();
+ // The provider was explicitly set, so check if the workspace has an agent model set.
+ if (this.invocation.workspace.agentModel)
+ return this.invocation.workspace.agentModel;
- // If all else fails - look at the provider default list
+ // Otherwise, we have no model to use - so guess a default model to use via the provider
+ // and it's system ENV params and if that fails - we return either a base model or null.
return this.providerDefault();
}
#providerSetupAndCheck() {
- this.provider = this.#workspace.agentProvider;
+ this.provider = this.#workspace.agentProvider ?? null;
this.model = this.#fetchModel();
+
+ if (!this.provider)
+ throw new Error("No valid provider found for the agent.");
this.log(`Start ${this.#invocationUUID}::${this.provider}:${this.model}`);
this.checkSetup();
}
diff --git a/server/utils/agents/index.js b/server/utils/agents/index.js
index ab86e51ff0e1fd19d5692c14c577b9177fdabd7c..98caea5cd37658cce1fb948062cef85b4a65889a 100644
--- a/server/utils/agents/index.js
+++ b/server/utils/agents/index.js
@@ -11,13 +11,6 @@ const ImportedPlugin = require("./imported");
class AgentHandler {
#invocationUUID;
#funcsToLoad = [];
- noProviderModelDefault = {
- azure: "OPEN_MODEL_PREF",
- lmstudio: "LMSTUDIO_MODEL_PREF",
- textgenwebui: null, // does not even use `model` in API req
- "generic-openai": "GENERIC_OPEN_AI_MODEL_PREF",
- bedrock: "AWS_BEDROCK_LLM_MODEL_PREFERENCE",
- };
invocation = null;
aibitat = null;
channel = null;
@@ -184,53 +177,70 @@ class AgentHandler {
}
}
+ /**
+ * Finds the default model for a given provider. If no default model is set for it's associated ENV then
+ * it will return a reasonable base model for the provider if one exists.
+ * @param {string} provider - The provider to find the default model for.
+ * @returns {string|null} The default model for the provider.
+ */
providerDefault(provider = this.provider) {
switch (provider) {
case "openai":
- return "gpt-4o";
+ return process.env.OPEN_MODEL_PREF ?? "gpt-4o";
case "anthropic":
- return "claude-3-sonnet-20240229";
+ return process.env.ANTHROPIC_MODEL_PREF ?? "claude-3-sonnet-20240229";
case "lmstudio":
- return "server-default";
+ return process.env.LMSTUDIO_MODEL_PREF ?? "server-default";
case "ollama":
- return "llama3:latest";
+ return process.env.OLLAMA_MODEL_PREF ?? "llama3:latest";
case "groq":
- return "llama3-70b-8192";
+ return process.env.GROQ_MODEL_PREF ?? "llama3-70b-8192";
case "togetherai":
- return "mistralai/Mixtral-8x7B-Instruct-v0.1";
+ return (
+ process.env.TOGETHER_AI_MODEL_PREF ??
+ "mistralai/Mixtral-8x7B-Instruct-v0.1"
+ );
case "azure":
- return "gpt-3.5-turbo";
- case "koboldcpp":
return null;
+ case "koboldcpp":
+ return process.env.KOBOLD_CPP_MODEL_PREF ?? null;
case "gemini":
- return "gemini-pro";
+ return process.env.GEMINI_MODEL_PREF ?? "gemini-pro";
case "localai":
- return null;
+ return process.env.LOCAL_AI_MODEL_PREF ?? null;
case "openrouter":
- return "openrouter/auto";
+ return process.env.OPENROUTER_MODEL_PREF ?? "openrouter/auto";
case "mistral":
- return "mistral-medium";
+ return process.env.MISTRAL_MODEL_PREF ?? "mistral-medium";
case "generic-openai":
- return null;
+ return process.env.GENERIC_OPEN_AI_MODEL_PREF ?? null;
case "perplexity":
- return "sonar-small-online";
+ return process.env.PERPLEXITY_MODEL_PREF ?? "sonar-small-online";
case "textgenwebui":
return null;
case "bedrock":
- return null;
+ return process.env.AWS_BEDROCK_LLM_MODEL_PREFERENCE ?? null;
case "fireworksai":
- return null;
+ return process.env.FIREWORKS_AI_LLM_MODEL_PREF ?? null;
case "deepseek":
- return "deepseek-chat";
+ return process.env.DEEPSEEK_MODEL_PREF ?? "deepseek-chat";
case "litellm":
- return null;
+ return process.env.LITE_LLM_MODEL_PREF ?? null;
case "apipie":
- return null;
+ return process.env.APIPIE_LLM_MODEL_PREF ?? null;
default:
- return "unknown";
+ return null;
}
}
+ /**
+ * Attempts to find a fallback provider and model to use if the workspace
+ * does not have an explicit `agentProvider` and `agentModel` set.
+ * 1. Fallback to the workspace `chatProvider` and `chatModel` if they exist.
+ * 2. Fallback to the system `LLM_PROVIDER` and try to load the the associated default model via ENV params or a base available model.
+ * 3. Otherwise, return null - will likely throw an error the user can act on.
+ * @returns {object|null} - An object with provider and model keys.
+ */
#getFallbackProvider() {
// First, fallback to the workspace chat provider and model if they exist
if (
@@ -262,7 +272,7 @@ class AgentHandler {
* If multi-model loading is supported, we use their agent model selection of the workspace
* If not supported, we attempt to fallback to the system provider value for the LLM preference
* and if that fails - we assume a reasonable base model to exist.
- * @returns {string} the model preference value to use in API calls
+ * @returns {string|null} the model preference value to use in API calls
*/
#fetchModel() {
// Provider was not explicitly set for workspace, so we are going to run our fallback logic
@@ -275,21 +285,11 @@ class AgentHandler {
}
// The provider was explicitly set, so check if the workspace has an agent model set.
- if (this.invocation.workspace.agentModel) {
+ if (this.invocation.workspace.agentModel)
return this.invocation.workspace.agentModel;
- }
- // If the provider we are using is not supported or does not support multi-model loading
- // then we use the default model for the provider.
- if (!Object.keys(this.noProviderModelDefault).includes(this.provider)) {
- return this.providerDefault();
- }
-
- // Load the model from the system environment variable for providers with no multi-model loading.
- const sysModelKey = this.noProviderModelDefault[this.provider];
- if (sysModelKey) return process.env[sysModelKey] ?? this.providerDefault();
-
- // Otherwise, we have no model to use - so guess a default model to use.
+ // Otherwise, we have no model to use - so guess a default model to use via the provider
+ // and it's system ENV params and if that fails - we return either a base model or null.
return this.providerDefault();
}
@@ -299,7 +299,6 @@ class AgentHandler {
if (!this.provider)
throw new Error("No valid provider found for the agent.");
-
this.log(`Start ${this.#invocationUUID}::${this.provider}:${this.model}`);
this.checkSetup();
}