Merge pull request #148 from dsanders11/feat/prompt-yaml-model-parameters

feat: support modelParameters in prompt.yaml files

__tests__/helpers-inference.test.ts

@@ -106,6 +106,8 @@ describe('helpers.ts - inference request building', () => {
         undefined,
         undefined,
         'gpt-4',
+        undefined,
+        undefined,
         100,
         'https://api.test.com',
         'test-token',
@@ -117,6 +119,8 @@ describe('helpers.ts - inference request building', () => {
           {role: 'user', content: 'User message'},
         ],
         modelName: 'gpt-4',
+        temperature: undefined,
+        topP: undefined,
         maxTokens: 100,
         endpoint: 'https://api.test.com',
         token: 'test-token',
@@ -136,6 +140,8 @@ describe('helpers.ts - inference request building', () => {
         'System prompt',
         'User prompt',
         'gpt-4',
+        undefined,
+        undefined,
         100,
         'https://api.test.com',
         'test-token',
@@ -147,6 +153,8 @@ describe('helpers.ts - inference request building', () => {
           {role: 'user', content: 'User prompt'},
         ],
         modelName: 'gpt-4',
+        temperature: undefined,
+        topP: undefined,
         maxTokens: 100,
         endpoint: 'https://api.test.com',
         token: 'test-token',

dist/index.js

@@ -49496,6 +49496,8 @@ async function simpleInference(request) {
         messages: request.messages,
         max_tokens: request.maxTokens,
         model: request.modelName,
+        temperature: request.temperature,
+        top_p: request.topP,
     };
     // Add response format if specified
     if (request.responseFormat) {
@@ -49530,6 +49532,8 @@ async function mcpInference(request, githubMcpClient) {
             messages: messages,
             max_tokens: request.maxTokens,
             model: request.modelName,
+            temperature: request.temperature,
+            top_p: request.topP,
         };
         // Add response format if specified (only on final iteration to avoid conflicts with tool calls)
         if (finalMessage && request.responseFormat) {
@@ -49685,12 +49689,14 @@ function buildResponseFormat(promptConfig) {
 /**
  * Build complete InferenceRequest from prompt config and inputs
  */
-function buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, maxTokens, endpoint, token) {
+function buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, temperature, topP, maxTokens, endpoint, token) {
     const messages = buildMessages(promptConfig, systemPrompt, prompt);
     const responseFormat = buildResponseFormat(promptConfig);
     return {
         messages,
         modelName,
+        temperature,
+        topP,
         maxTokens,
         endpoint,
         token,
@@ -52640,7 +52646,10 @@ async function run() {
         }
         // Get common parameters
         const modelName = promptConfig?.model || coreExports.getInput('model');
-        const maxTokens = parseInt(coreExports.getInput('max-tokens'), 10);
+        let maxTokens = promptConfig?.modelParameters?.maxTokens ?? coreExports.getInput('max-tokens');
+        if (typeof maxTokens === 'string') {
+            maxTokens = parseInt(maxTokens, 10);
+        }
         const token = process.env['GITHUB_TOKEN'] || coreExports.getInput('token');
         if (token === undefined) {
             throw new Error('GITHUB_TOKEN is not set');
@@ -52649,7 +52658,7 @@ async function run() {
         const githubMcpToken = coreExports.getInput('github-mcp-token') || token;
         const endpoint = coreExports.getInput('endpoint');
         // Build the inference request with pre-processed messages and response format
-        const inferenceRequest = buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, maxTokens, endpoint, token);
+        const inferenceRequest = buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, promptConfig?.modelParameters?.temperature, promptConfig?.modelParameters?.topP, maxTokens, endpoint, token);
         const enableMcp = coreExports.getBooleanInput('enable-github-mcp') || false;
         let modelResponse = null;
         if (enableMcp) {

src/helpers.ts

@@ -82,6 +82,8 @@ export function buildInferenceRequest(
   systemPrompt: string | undefined,
   prompt: string | undefined,
   modelName: string,
+  temperature: number | undefined,
+  topP: number | undefined,
   maxTokens: number,
   endpoint: string,
   token: string,
@@ -92,6 +94,8 @@ export function buildInferenceRequest(
   return {
     messages,
     modelName,
+    temperature,
+    topP,
     maxTokens,
     endpoint,
     token,

src/inference.ts

@@ -15,6 +15,8 @@ export interface InferenceRequest {
   maxTokens: number
   endpoint: string
   token: string
+  temperature?: number
+  topP?: number
   responseFormat?: {type: 'json_schema'; json_schema: unknown} // Processed response format for the API
 }
 
@@ -45,6 +47,8 @@ export async function simpleInference(request: InferenceRequest): Promise<string
     messages: request.messages as OpenAI.Chat.Completions.ChatCompletionMessageParam[],
     max_tokens: request.maxTokens,
     model: request.modelName,
+    temperature: request.temperature,
+    top_p: request.topP,
   }
 
   // Add response format if specified
@@ -90,6 +94,8 @@ export async function mcpInference(
       messages: messages as OpenAI.Chat.Completions.ChatCompletionMessageParam[],
       max_tokens: request.maxTokens,
       model: request.modelName,
+      temperature: request.temperature,
+      top_p: request.topP,
     }
 
     // Add response format if specified (only on final iteration to avoid conflicts with tool calls)

src/main.ts

@@ -53,7 +53,11 @@ export async function run(): Promise<void> {
 
     // Get common parameters
     const modelName = promptConfig?.model || core.getInput('model')
-    const maxTokens = parseInt(core.getInput('max-tokens'), 10)
+    let maxTokens = promptConfig?.modelParameters?.maxTokens ?? core.getInput('max-tokens')
+
+    if (typeof maxTokens === 'string') {
+      maxTokens = parseInt(maxTokens, 10)
+    }
 
     const token = process.env['GITHUB_TOKEN'] || core.getInput('token')
     if (token === undefined) {
@@ -71,6 +75,8 @@ export async function run(): Promise<void> {
       systemPrompt,
       prompt,
       modelName,
+      promptConfig?.modelParameters?.temperature,
+      promptConfig?.modelParameters?.topP,
       maxTokens,
       endpoint,
       token,

src/prompt.ts

@@ -7,9 +7,16 @@ export interface PromptMessage {
   content: string
 }
 
+export interface ModelParameters {
+  maxTokens?: number
+  temperature?: number
+  topP?: number
+}
+
 export interface PromptConfig {
   messages: PromptMessage[]
   model?: string
+  modelParameters?: ModelParameters
   responseFormat?: 'text' | 'json_schema'
   jsonSchema?: string
 }