feat: use entire file in completion prompt

.env.example

@@ -1,4 +1,5 @@
-OPENAI_MODEL=gpt-3.5-turbo
+OPENAI_MODEL=gpt-3.5-turbo-16k
+OPENAI_MAX_TOKENS=8096
 OPENAI_API_KEY=
 LOG_FILE=/app/helix-gpt.log
 OPENAI_CONTEXT=

README.md

@@ -16,8 +16,9 @@ You can configure helix-gpt by exposing either the environment variables below,
 
 Environment vars
 ```
-OPENAI_MODEL=gpt-3.5-turbo # Optional
+OPENAI_MODEL=gpt-3.5-turbo-16k # Optional
 OPENAI_API_KEY=123 # required
+OPENAI_MAX_TOKENS=7000 # optional
 LOG_FILE=/app/debug-helix-gpt.log # Optional
 OPENAI_CONTEXT="A terrible code completion assistant" # Optional
 ```

src/app.ts

@@ -67,21 +67,25 @@ const main = async () => {
     }
 
     log("calling completion event", contentVersion, "<", lastContentVersion)
-    const content = await getContent(contents, request.params.position.line, request.params.position.character)
+    const { lastCharacter, lastLine, templatedContent } = await getContent(contents, request.params.position.line, request.params.position.character)
 
-    if (!triggerCharacters.includes(content.slice(-1))) {
-      log("skipping", content.slice(-1), "not in", triggerCharacters)
+    if (!triggerCharacters.includes(lastCharacter)) {
+      log("skipping", lastCharacter, "not in", triggerCharacters)
       skip()
       return
     }
 
-    const hints = await getHints(content, language)
+    const hints = await getHints(templatedContent, language)
 
     log("sending completion", JSON.stringify({
-      content, hints
+      templatedContent, hints
     }))
 
     const items = hints?.map((i) => {
+      if (i.startsWith(lastLine.trim())) {
+        i = i.slice(lastLine.trim().length)
+      }
+
       const lines = i.split('\n')
       const cleanLine = request.params.position.line + lines.length - 1
       let cleanCharacter = lines.slice(-1)[0].length

src/config.ts

@@ -14,7 +14,11 @@ const { values } = parseArgs({
     },
     openaiModel: {
       type: 'string',
-      default: Bun.env.OPENAI_MODEL ?? "gpt-3.5-turbo"
+      default: Bun.env.OPENAI_MODEL ?? "gpt-3.5-turbo-16k"
+    },
+    openaiMaxTokens: {
+      type: 'string',
+      default: Bun.env.OPENAI_MAX_TOKENS ?? "7000"
     },
     logFile: {
       type: 'string',

src/constants.ts

@@ -1,34 +1,16 @@
-export const context = `As a language model optimized for code completion in the language <languageId>, follow these principles:
-
-- Analyze the given code snippet to identify what type of construct it represents (e.g., function, loop, conditional statement).
-- Discern any identifiable coding patterns or conventions in the provided snippet to maintain consistency in style and structure.
-- Complete only the immediate section of code that is being worked on without expanding beyond its scope.
-- Avoid adding comments or annotations within your response since they are not requested.
-- Refrain from repeating any part of the original request's code in your output; focus solely on generating new content that logically and syntactically follows from it.`
-
+export const context = `Continue the input code from the language <languageId> at the <BEGIN_COMPLETION> marker. Only respond with code.`
 export const examples = [
   {
     role: "user",
-    content: "import example from \"./model\"\n\nconst testing = 123\nfunction randomString(",
+    content: `function randomInt(<BEGIN_COMPLETION>`
   },
   {
     role: "assistant",
-    content: `length: number): string {
-    let result = '';
-    const characters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789';
-    for (let i = 0; i < length; i++) {
-       result += characters.charAt(Math.floor(Math.random() * characters.length));
-    }
-    return result;
+    content: `min: number, max: number): number {
+    min = Math.ceil(min);
+    max = Math.floor(max);
+    return Math.floor(Math.random() * (max - min + 1)) + min;
 }`
-  },
-  {
-    role: "user",
-    content: "import test from \"util\"\n\nconst alphabet ",
-  },
-  {
-    role: "assistant",
-    content: " = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ';"
   }
 ]
 

src/openai.ts

@@ -17,11 +17,12 @@ export const getHints = async (contents: string, languageId: string, suggestions
 
   const body = {
     "model": config.openaiModel,
-    "max_tokens": 2048,
+    "max_tokens": parseInt(config.openaiMaxTokens as string),
     "n": suggestions,
-    "temperature": 0.7,
-    "top_p": 1,
-    "frequency_penalty": 0,
+    "temperature": 1,
+    "top_p": 0.7,
+    "frequency_penalty": 1,
+    presence_penalty: 2,
     messages
   }
 
@@ -37,6 +38,7 @@ export const getHints = async (contents: string, languageId: string, suggestions
   })
 
   if (!response.ok) {
+    log("openai error", response.status, await response.text())
     throw new Error("openai request failed with code: " + response.status)
   }
 

src/utils.ts

@@ -4,7 +4,13 @@ import config from "./config"
 export const getContent = async (contents: string, line: number, column: number) => {
   const lines = contents.split('\n').slice(0, line + 1)
   lines[lines.length - 1] = lines[lines.length - 1].split('').slice(0, column).join('')
-  return lines.join('\n')
+  const lastLine = lines[lines.length - 1]
+  const contentBefore = lines.join('\n')
+  const contentAfter = contents.split('\n').slice(line + 1).join('\n')
+  const lastCharacter = contentBefore.slice(-1)
+  const templatedContent = `${contentBefore}<BEGIN_COMPLETION>\n${contentAfter}`
+
+  return { contentBefore, contentAfter, lastCharacter, templatedContent, lastLine }
 }
 
 export const log = (...args: any) => {