Pass more config options through to LLM providers

DeabLabs · Sep 8, 2024 · de59766 · de59766
1 parent 2de6b85
commit de59766
Showing 1 changed file with 36 additions and 33 deletions.
diff --git a/packages/cannoli-core/src/providers.ts b/packages/cannoli-core/src/providers.ts
@@ -46,6 +46,8 @@ export const GenericModelConfigSchema = z.object({
 	microstat: z.boolean().optional(),
 	microstat_eta: z.coerce.number().optional(),
 	microstat_tau: z.coerce.number().optional(),
+	max_tokens: z.coerce.number().optional(),
+	user: z.string().optional(),
 	num_ctx: z.coerce.number().optional(),
 	num_gqa: z.coerce.number().optional(),
 	num_gpu: z.coerce.number().optional(),
@@ -80,35 +82,6 @@ export type GenericCompletionResponse = {
 	function_call?: ChatCompletionAssistantMessageParam.FunctionCall;
 };
 
-// @deprecated
-export const makeSampleConfig = (): GenericModelConfig => ({
-	apiKey: undefined,
-	baseURL: undefined,
-	model: "",
-	frequency_penalty: undefined,
-	presence_penalty: undefined,
-	stop: undefined,
-	function_call: undefined,
-	functions: undefined,
-	temperature: undefined,
-	top_p: undefined,
-	role: "user" || "assistant" || "system",
-	provider: undefined,
-	microstat: undefined,
-	microstat_eta: undefined,
-	microstat_tau: undefined,
-	num_ctx: undefined,
-	num_gqa: undefined,
-	num_gpu: undefined,
-	num_thread: undefined,
-	repeat_last_n: undefined,
-	repeat_penalty: undefined,
-	seed: undefined,
-	tfs_z: undefined,
-	num_predict: undefined,
-	top_k: undefined,
-});
-
 export type GetDefaultsByProvider = (provider: SupportedProviders) => GenericModelConfig | undefined;
 
 export type LangchainMessages = ReturnType<typeof LLMProvider.convertMessages>;
@@ -163,10 +136,6 @@ export class LLMProvider {
 		return defaults;
 	}
 
-	getSampleConfig() {
-		return makeSampleConfig();
-	}
-
 	getMergedConfig = (args?: Partial<{
 		configOverrides: GenericModelConfig;
 		provider: SupportedProviders;
@@ -202,6 +171,15 @@ export class LLMProvider {
 					apiKey: config.apiKey,
 					model: config.model,
 					temperature: config.temperature,
+					topP: config.top_p,
+					frequencyPenalty: config.frequency_penalty,
+					presencePenalty: config.presence_penalty,
+					stop: config.stop?.split(","),
+					maxTokens: config.max_tokens,
+					user: config.user,
+					// beta openai feature
+					// @ts-expect-error
+					seed: config.seed,
 					maxRetries: 3,
 					configuration: {
 						baseURL: url,
@@ -218,6 +196,15 @@ export class LLMProvider {
 					azureOpenAIApiInstanceName: config.azureOpenAIApiInstanceName,
 					azureOpenAIApiVersion: config.azureOpenAIApiVersion,
 					azureOpenAIBasePath: url,
+					user: config.user,
+					maxTokens: config.max_tokens,
+					// beta openai feature
+					// @ts-expect-error
+					seed: config.seed,
+					topP: config.top_p,
+					frequencyPenalty: config.frequency_penalty,
+					presencePenalty: config.presence_penalty,
+					stop: config.stop?.split(","),
 					maxRetries: 3,
 					configuration: {
 						baseURL: url,
@@ -230,20 +217,31 @@ export class LLMProvider {
 						baseUrl: url,
 						model: config.model,
 						temperature: config.temperature,
+						topP: config.top_p,
+						frequencyPenalty: config.frequency_penalty,
+						presencePenalty: config.presence_penalty,
+						stop: config.stop?.split(","),
 					});
 				}
 
 				return new ChatOllama({
 					baseUrl: url,
 					model: config.model,
 					temperature: config.temperature,
+					topP: config.top_p,
+					frequencyPenalty: config.frequency_penalty,
+					presencePenalty: config.presence_penalty,
+					stop: config.stop?.split(","),
 				});
 			case "gemini":
 				return new ChatGoogleGenerativeAI({
 					maxRetries: 3,
 					model: config.model,
 					apiKey: config.apiKey,
 					temperature: config.temperature,
+					maxOutputTokens: config.max_tokens,
+					topP: config.top_p,
+					stopSequences: config.stop?.split(","),
 				});
 			case "anthropic":
 				return new ChatAnthropic({
@@ -252,6 +250,10 @@ export class LLMProvider {
 					temperature: config.temperature,
 					maxRetries: 0,
 					anthropicApiUrl: url,
+					maxTokens: config.max_tokens,
+					topP: config.top_p,
+					stopSequences: config.stop?.split(","),
+					topK: config.top_k,
 					clientOptions: {
 						defaultHeaders: {
 							"anthropic-dangerous-direct-browser-access": "true",
@@ -263,6 +265,7 @@ export class LLMProvider {
 					apiKey: config.apiKey,
 					model: config.model,
 					temperature: config.temperature,
+					stopSequences: config.stop?.split(","),
 					maxRetries: 3,
 				});
 			default: