Phase 3: backends - llama.cpp (local) and cloud (OpenAI/DeepSeek/Anthropic)

app/src/main/java/io/github/abrarul/modularai/engine/Engine.kt

@@ -0,0 +1,92 @@
+package io.github.abrarul.modularai.engine
+
+import org.json.JSONArray
+import org.json.JSONObject
+import java.net.HttpURLConnection
+import java.net.URL
+import java.nio.ByteBuffer
+import java.nio.CharBuffer
+import java.nio.charset.CodingErrorAction
+
+interface Backend {
+    /** Streams text through [onText]; return false from it to stop. */
+    fun generate(prompt: String, maxTokens: Int, onText: (String) -> Boolean)
+}
+
+/** Called from native code. Keep the name and signature in sync with llama_jni.cpp and proguard. */
+interface TokenSink {
+    fun onToken(bytes: ByteArray): Boolean
+}
+
+class LlamaEngine : Backend {
+    private var handle = 0L
+
+    private external fun nativeLoad(path: String, nCtx: Int, threads: Int): Long
+    private external fun nativeGenerate(h: Long, prompt: String, maxTokens: Int, temp: Float, sink: TokenSink)
+    private external fun nativeFree(h: Long)
+
+    val loaded get() = handle != 0L
+
+    fun load(path: String, nCtx: Int, threads: Int = Runtime.getRuntime().availableProcessors().coerceAtMost(6)) {
+        unload()
+        handle = nativeLoad(path, nCtx, threads)
+        require(handle != 0L) { "Could not load model (unsupported file, or not enough RAM)" }
+    }
+
+    fun unload() { if (handle != 0L) { nativeFree(handle); handle = 0L } }
+
+    override fun generate(prompt: String, maxTokens: Int, onText: (String) -> Boolean) {
+        check(handle != 0L) { "No model loaded" }
+        val dec = Charsets.UTF_8.newDecoder()
+            .onMalformedInput(CodingErrorAction.REPLACE).onUnmappableCharacter(CodingErrorAction.REPLACE)
+        var pending = ByteArray(0)
+        nativeGenerate(handle, prompt, maxTokens, 0.7f, object : TokenSink {
+            override fun onToken(bytes: ByteArray): Boolean {
+                // A token may end mid-character; decode only complete UTF-8 sequences.
+                val bb = ByteBuffer.wrap(pending + bytes)
+                val cb = CharBuffer.allocate(bb.remaining() + 4)
+                dec.decode(bb, cb, false)
+                pending = ByteArray(bb.remaining()).also { bb.get(it) }
+                cb.flip()
+                return if (cb.isEmpty()) true else onText(cb.toString())
+            }
+        })
+    }
+
+    companion object { init { System.loadLibrary("modularai_llama") } }
+}
+
+/** Non-streaming cloud fallback. [provider]: openai | deepseek | anthropic. */
+class CloudBackend(private val provider: String, private val key: String, private val model: String) : Backend {
+
+    override fun generate(prompt: String, maxTokens: Int, onText: (String) -> Boolean) {
+        val msgs = JSONArray().put(JSONObject().put("role", "user").put("content", prompt))
+        val anthropic = provider == "anthropic"
+        val url = when (provider) {
+            "anthropic" -> "https://api.anthropic.com/v1/messages"
+            "deepseek" -> "https://api.deepseek.com/chat/completions"
+            "openai" -> "https://api.openai.com/v1/chat/completions"
+            else -> error("Unknown provider: $provider (openai | deepseek | anthropic)")
+        }
+        val body = JSONObject().put("model", model).put("max_tokens", maxTokens).put("messages", msgs)
+        val c = URL(url).openConnection() as HttpURLConnection
+        c.requestMethod = "POST"
+        c.doOutput = true
+        c.connectTimeout = 20_000
+        c.readTimeout = 120_000
+        c.setRequestProperty("Content-Type", "application/json")
+        if (anthropic) {
+            c.setRequestProperty("x-api-key", key)
+            c.setRequestProperty("anthropic-version", "2023-06-01")
+        } else c.setRequestProperty("Authorization", "Bearer $key")
+        c.outputStream.use { it.write(body.toString().toByteArray()) }
+
+        val ok = c.responseCode in 200..299
+        val text = (if (ok) c.inputStream else c.errorStream).bufferedReader().readText()
+        if (!ok) error("HTTP ${c.responseCode}: ${text.take(300)}")
+        val j = JSONObject(text)
+        val out = if (anthropic) j.getJSONArray("content").getJSONObject(0).getString("text")
+        else j.getJSONArray("choices").getJSONObject(0).getJSONObject("message").getString("content")
+        onText(out)
+    }
+}