refactor: convert settings.js to .mjs, class-ify cache.mjs, disk load in CachedInferenceEngine [opencode:big-pickle]

Author: geoffroy-noel-ddh

app/cached-inference-engine.mjs

@@ -7,11 +7,26 @@ export default class CachedInferenceEngine extends InferenceEngine {
   }
 
   async loadCache(url) {
-    const response = await fetch(url)
-    if (!response.ok) {
-      throw new Error(`HTTP error! status: ${response.status}`)
+    let ret = null
+    const isNode = typeof process !== 'undefined' && process.versions != null && process.versions.node != null
+    const urlStr = typeof url === 'string' ? url : url.href
+    if (isNode) {
+      if (!urlStr.startsWith('http:') && !urlStr.startsWith('https:')) {
+        const { readFileSync } = await import('node:fs')
+        const { fileURLToPath } = await import('node:url')
+        const filePath = urlStr.startsWith('file:') ? fileURLToPath(urlStr) : urlStr
+        ret = JSON.parse(readFileSync(filePath, 'utf-8'))
+      }
+    }
+    if (ret === null) {
+      const response = await fetch(url)
+      if (!response.ok) {
+        throw new Error(`HTTP error! status: ${response.status}`)
+      }
+      ret = await response.json()
     }
-    this.cache = await response.json()
+    this.cache = ret
+    return ret
   }
 
   getCache() {

app/nlpui.html

@@ -159,7 +159,7 @@ <h2 class="title is-3">Defendant's statement (with LLM's highlights)</h2>
         </div>
       </div>
     </section>
-    <script src="settings.js"></script>
+    <script type="module" src="settings.mjs"></script>
     <script type="module" src="nlpui.js"></script>
   </body>
 </html>

app/poc3.html

@@ -101,7 +101,7 @@ <h2 class="title is-3">Case</h2>
         </div>
       </div>
     </section>
-    <script src="settings.js"></script>
+    <script type="module" src="settings.mjs"></script>
     <script type="module" src="poc3.js"></script>
   </body>
 </html>

app/poc3.js

@@ -5,7 +5,7 @@ DONE case selection
 DONE show case
 DONE show penalty
 DONE table
-M draw LLM responses from cache
+DONE draw LLM responses from cache
 S explaination
 C highlights
 C summary 

app/settings.js app/settings.mjsapp/settings.js renamed to app/settings.mjs

@@ -166,4 +166,6 @@ const SETTINGS = {
   },
 }
 
-window.SETTINGS = SETTINGS
+if (typeof window !== 'undefined') {
+  window.SETTINGS = SETTINGS
+}

tools/cache.mjs

@@ -1,116 +1,111 @@
 #!/usr/bin/env node
 import { readFileSync, writeFileSync } from 'node:fs'
-import { DEFAULT_TEMPLATE_QUESTIONNAIRE as TEMPLATE } from '../app/settings.js'
+import { DEFAULT_TEMPLATE_QUESTIONNAIRE as TEMPLATE } from '../app/settings.mjs'
+import CachedInferenceEngine from '../app/cached-inference-engine.mjs'
 
-const CACHE_FILE = new URL('../app/data/responses.json', import.meta.url)
 const APP_DATA_FILE = new URL('../app/data/app-data.json', import.meta.url)
 
-function load() {
-  return JSON.parse(readFileSync(CACHE_FILE, 'utf-8'))
-}
-
-function actionLs() {
-  const data = load()
+class ResponseCacheManager {
 
-  for (const [key, entry] of Object.entries(data)) {
-    console.log(`${key}\t${entry.model}\t${entry.response.length}`)
+  constructor() {
+    this.engine = new CachedInferenceEngine({
+      serviceUrl: process.env.AJ_LLM_API,
+      model: process.env.AJ_MODEL
+    })
   }
-}
 
-function actionModels() {
-  const data = load()
-  const counts = {}
-
-  for (const entry of Object.values(data)) {
-    counts[entry.model] = (counts[entry.model] || 0) + 1
+  async loadCache() {
+    await this.engine.loadCache('../app/data/responses.json')
   }
 
-  for (const [model, count] of Object.entries(counts)) {
-    console.log(`${model}\t${count}`)
+  getCacheEntries() {
+    return this.engine.getCache()
   }
-}
-
-function hash(string) {
-  let h = 0
-  for (const char of string) {
-    h = (h << 5) - h + char.charCodeAt(0)
-    h |= 0
-  }
-  return h
-}
-
-function buildPrompt(template, variables) {
-  let ret = template
-  for (const [k, v] of Object.entries(variables)) {
-    ret = ret.replace(`{${k}}`, v)
-  }
-  return ret
-}
 
-async function actionFetch() {
-  const serviceUrl = process.env.AJ_LLM_API
-  const model = process.env.AJ_MODEL
+  actionLs() {
+    const entries = this.getCacheEntries()
 
-  if (!serviceUrl || !model) {
-    console.error('Set AJ_LLM_API and AJ_MODEL environment variables')
-    process.exit(1)
+    for (const [key, entry] of Object.entries(entries)) {
+      console.log(`${key}\t${entry.model}\t${entry.response.length}`)
+    }
   }
 
-  const cache = load()
-  const appData = JSON.parse(readFileSync(APP_DATA_FILE, 'utf-8'))
-
-  for (const [caseKey, caseData] of Object.entries(appData.cases)) {
-    for (const question of appData.questions) {
-      const prompt = buildPrompt(TEMPLATE, {
-        STATEMENT: caseData.statement,
-        QUESTION: question.text,
-      })
+  actionModels() {
+    const entries = this.getCacheEntries()
+    const counts = {}
 
-      const key = hash(`${model}-${prompt}`)
+    for (const entry of Object.values(entries)) {
+      counts[entry.model] = (counts[entry.model] || 0) + 1
+    }
 
-      if (cache[key]) {
-        continue
-      }
+    for (const [model, count] of Object.entries(counts)) {
+      console.log(`${model}\t${count}`)
+    }
+  }
 
-      const url = serviceUrl.replace(/\/+$/, '') + '/chat/completions'
-      const res = await fetch(url, {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-          model,
-          messages: [{ role: 'user', content: prompt }],
-          stream: false,
-          max_tokens: 4000,
-        }),
-      })
-
-      if (!res.ok) {
-        console.error(`API error (${res.status}) for ${caseKey} / ${question.text}`)
-        continue
+  async actionFetch() {
+    const cache = this.engine.getCache()
+    const appData = JSON.parse(readFileSync(APP_DATA_FILE, 'utf-8'))
+
+    for (const [caseKey, caseData] of Object.entries(appData.cases)) {
+      for (const question of appData.questions) {
+        const prompt = this.engine.getPromptFromTemplate(TEMPLATE, {
+          STATEMENT: caseData.statement,
+          QUESTION: question.text,
+        })
+
+        const key = CachedInferenceEngine.hash(`${model}-${prompt}`)
+
+        if (cache[key]) {
+          continue
+        }
+
+        const url = serviceUrl.replace(/\/+$/, '') + '/chat/completions'
+        const res = await fetch(url, {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({
+            model,
+            messages: [{ role: 'user', content: prompt }],
+            stream: false,
+            max_tokens: 4000,
+          }),
+        })
+
+        if (!res.ok) {
+          console.error(`API error (${res.status}) for ${caseKey} / ${question.text}`)
+          continue
+        }
+
+        const data = await res.json()
+        const response = data?.choices?.[0]?.message?.content
+
+        if (response) {
+          cache[key] = { response, model, hash: key }
+          console.log(`cached ${caseKey} / ${question.text}`)
+        }
       }
+    }
+  }
 
-      const data = await res.json()
-      const response = data?.choices?.[0]?.message?.content
-
-      if (response) {
-        cache[key] = { response, model, hash: key }
-        console.log(`cached ${caseKey} / ${question.text}`)
+  async runAction(action) {
+    await this.loadCache()
+
+    if (action === 'ls') {
+      this.actionLs()
+    } else if (action === 'models') {
+      this.actionModels()
+    } else if (action === 'fetch') {
+      this.actionFetch()
+    } else {
+      if (action && action !== 'help') {
+        console.error('Unknown action:', action)
       }
-
-      writeFileSync(CACHE_FILE, JSON.stringify(cache, null, 2))
+      console.error('Usage: node cache.mjs <ls|models|fetch>')
+      process.exit(1)
     }
   }
 }
 
-const action = process.argv[2]
-if (action === 'ls') {
-  actionLs()
-} else if (action === 'models') {
-  actionModels()
-} else if (action === 'fetch') {
-  actionFetch()
-} else {
-  console.error('Unknown action:', action)
-  console.error('Usage: node cache.mjs <ls|models|fetch>')
-  process.exit(1)
-}
+const app = new ResponseCacheManager()
+await app.runAction(process.argv[2])