mmanningau/alice
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Have settled on testing and using the Qwen2.5:1.5b model as it plays much nicer with the JSON requirements - way better than the Gemma models did. Have refined the calls via LlmManager.kt and MediaPipeAdapter.kt and tested using battery, ping and a new weather skill. Half way through updating further to allow for SMS reading and updating the SkillManager.kt again to accommodate this with a two-pass bridge for python calling.

Browse Source
This commit is contained in:
mmanningau
2026-03-05 15:13:56 +11:00
parent 8ba9cb7a62
commit 2f9b00ae29
5 changed files with 96 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
app/src/main/AndroidManifest.xml
View File

@@ -6,6 +6,10 @@
<uses-permission android:name="android.permission.BIND_ACCESSIBILITY_SERVICE" />
<uses-permission android:name="android.permission.READ_SMS"/>
<uses-permission android:name="android.permission.RECEIVE_SMS"/> <!-- for new messages -->
<application
android:name="AliceApp"
android:usesCleartextTraffic="true"

41
app/src/main/java/net/mmanningau/alice/LlmManager.kt
View File

@@ -141,44 +141,63 @@ object LlmManager {
if (currentMode == "Local" && chatModel == null) return "System: Local engine is selected but not properly initialized or unsupported file format."
val currentModel = chatModel ?: return "Error: LLM engine not initialized."
// If the history size is 1, it means only the System prompt exists. This is the first message!
if (chatHistory.size == 1) {
// Take the first 25 characters. If it's longer, add "..."
val previewLength = 25
val newTitle = if (userText.length > previewLength) {
userText.take(previewLength).trim() + "..."
} else {
userText
}
// Update the database instantly
val newTitle = if (userText.length > previewLength) userText.take(previewLength).trim() + "..." else userText
chatDao?.updateThreadTitle(currentThreadId, newTitle)
}
// 1. Save user message to DB and Memory
chatDao?.insertMessage(ChatMessageEntity(threadId = currentThreadId, text = userText, isUser = true))
chatHistory.add(UserMessage(userText))
val toolSpecs = SkillManager.loadSkills()
// --- LOOP CONTROL CONSTANTS ---
val MAX_TOOL_ITERATIONS = 5
var toolIterations = 0
val executedToolSignatures = mutableSetOf<String>() // Tracks name+args pairs to catch spin loops
var response = currentModel.generate(chatHistory, toolSpecs)
var aiMessage: AiMessage = response.content()
chatHistory.add(aiMessage)
while (aiMessage.hasToolExecutionRequests()) {
// --- GUARD 1: Hard iteration cap ---
if (toolIterations >= MAX_TOOL_ITERATIONS) {
Log.w("AliceEngine", "Tool loop cap reached after $MAX_TOOL_ITERATIONS iterations. Breaking.")
val fallbackText = "I've reached the maximum number of steps trying to complete this task. Here's what I found so far."
chatDao?.insertMessage(ChatMessageEntity(threadId = currentThreadId, text = fallbackText, isUser = false))
return fallbackText
}
for (request in aiMessage.toolExecutionRequests()) {
val toolName = request.name()
val arguments = request.arguments()
// --- GUARD 2: Duplicate call detection ---
val signature = "$toolName::$arguments"
if (executedToolSignatures.contains(signature)) {
Log.w("AliceEngine", "Duplicate tool call detected for '$toolName'. Breaking loop.")
val fallbackText = "I seem to be going in circles with the '$toolName' tool. Let me stop and give you what I have."
chatDao?.insertMessage(ChatMessageEntity(threadId = currentThreadId, text = fallbackText, isUser = false))
return fallbackText
}
executedToolSignatures.add(signature)
val toolResult = SkillManager.executeSkill(toolName, arguments)
Log.d("AliceSkill", "TOOL_RESULT from [$toolName]: $toolResult")
chatHistory.add(ToolExecutionResultMessage(request.id(), toolName, toolResult))
}
toolIterations++
response = currentModel.generate(chatHistory, toolSpecs)
aiMessage = response.content()
chatHistory.add(aiMessage)
}
// 2. Save final AI message to DB
chatDao?.insertMessage(ChatMessageEntity(threadId = currentThreadId, text = aiMessage.text(), isUser = false))
return aiMessage.text()
}
}

6
app/src/main/java/net/mmanningau/alice/MediaPipeAdapter.kt
View File

@@ -35,7 +35,7 @@ class MediaPipeAdapter(
// THE PUSH: Aggressively demand the Adreno GPU
val gpuOptions = LlmInference.LlmInferenceOptions.builder()
.setModelPath(modelPath)
.setMaxTokens(4096)
.setMaxTokens(1200)
.setPreferredBackend(LlmInference.Backend.GPU)
.build()
engine = LlmInference.createFromOptions(context, gpuOptions)
@@ -50,7 +50,7 @@ class MediaPipeAdapter(
val cpuOptions = LlmInference.LlmInferenceOptions.builder()
.setModelPath(modelPath)
.setMaxTokens(4096)
.setMaxTokens(1200)
.setPreferredBackend(LlmInference.Backend.CPU)
.build()
engine = LlmInference.createFromOptions(context, cpuOptions)
@@ -105,7 +105,7 @@ class MediaPipeAdapter(
}
}
is ToolExecutionResultMessage -> {
promptBuilder.append("<start_of_turn>user\n[SYSTEM DATA VIA TOOL '${message.toolName()}']: ${message.text()}\nUse this real data to answer the previous question.<end_of_turn>\n")
promptBuilder.append("<start_of_turn>user\n[TOOL RESULT: ${message.toolName()}]\n${message.text()}\n\nIMPORTANT: The above is raw data from a tool. Do NOT repeat it verbatim. You must now write a naturally worded and concise response to the user's original question using this data. Summarise it concisely as Alice their helpful AI assistant.<end_of_turn>\n")
}
is AiMessage -> {
if (message.hasToolExecutionRequests()) {

10
app/src/main/java/net/mmanningau/alice/ModelRegistry.kt
View File

@@ -64,12 +64,12 @@ object ModelRegistry {
downloadUrl = "https://huggingface.co/google/gemma-3n-E2B-it-litert-lm/resolve/main/gemma-3n-E2B-it-int4.litertlm"
),
LocalModel(
id = "Gemma3-1B-IT_multi-prefill-seq_q8_ekv4096",
name = "Gemma 3 (1B) Prefill",
sizeMb = 3390,
id = "Qwen2.5-1.5B-Instruct_seq128_q8_ekv1280",
name = "Qwen2.5-1.5B",
sizeMb = 1570,
description = "A highly optimised and fine tuned model for agentic tasks and function calling.",
fileName = "Gemma3-1B-IT_multi-prefill-seq_q8_ekv4096.task",
downloadUrl = "https://huggingface.co/litert-community/Gemma3-1B-IT/resolve/main/Gemma3-1B-IT_multi-prefill-seq_q8_ekv4096.task"
fileName = "Qwen2.5-1.5B-Instruct_seq128_q8_ekv1280.task",
downloadUrl = "https://huggingface.co/litert-community/Qwen2.5-1.5B-Instruct/resolve/main/Qwen2.5-1.5B-Instruct_seq128_q8_ekv1280.task"
)
)

66
app/src/main/java/net/mmanningau/alice/SkillManager.kt
View File

@@ -1,10 +1,12 @@
package net.mmanningau.alice
import android.content.Context
import android.net.Uri
import android.util.Log
import com.chaquo.python.Python
import dev.langchain4j.agent.tool.JsonSchemaProperty
import dev.langchain4j.agent.tool.ToolSpecification
import org.json.JSONArray
import org.json.JSONObject
import java.io.File
@@ -12,7 +14,11 @@ object SkillManager {
var skillsDirectory: File? = null
private set
// *** ADDED: store context for ContentResolver access
private var appContext: Context? = null
fun initialize(context: Context) {
appContext = context.applicationContext // *** ADDED
val baseDir = context.getExternalFilesDir(null)
val skillsDir = File(baseDir, "Skills")
@@ -25,7 +31,7 @@ object SkillManager {
fun updateDirectory(newPath: String) {
val newDir = File(newPath)
if (!newDir.exists()) {
newDir.mkdirs() // Create it if the user typed a new path
newDir.mkdirs()
}
skillsDirectory = newDir
Log.i("AliceSkills", "Skills directory updated to: ${newDir.absolutePath}")
@@ -48,13 +54,12 @@ object SkillManager {
.name(name)
.description(description)
// Parse the expected parameters so the LLM knows what to extract
val parameters = json.optJSONObject("parameters")
val properties = parameters?.optJSONObject("properties")
properties?.keys()?.forEach { key ->
val prop = properties.getJSONObject(key)
val type = prop.getString("type") // e.g., "string"
val type = prop.getString("type")
val desc = prop.optString("description", "")
builder.addParameter(
@@ -85,23 +90,60 @@ object SkillManager {
val py = Python.getInstance()
val builtins = py.builtins
// We create an isolated dictionary for the script to run in.
// This allows you to edit the Python files and have them hot-reload instantly!
val globals = py.getModule("builtins").callAttr("dict")
// Execute the raw script text
builtins.callAttr("exec", scriptFile.readText(), globals)
// Find the 'execute' function we mandated in our python script
val executeFunc = globals.callAttr("get", "execute")
if (executeFunc == null) return "Error: Python script missing 'def execute(args_json):' function."
if (executeFunc == null) return "Error: Python script missing 'def execute(args):' function."
// Call it and return the string!
executeFunc.call(argumentsJson).toString()
// First call to Python
var result = executeFunc.call(argumentsJson).toString()
// *** ADDED: Two-pass bridge for skills that need Android ContentResolver
if (result.startsWith("BRIDGE_REQUEST:")) {
val ctx = appContext
if (ctx == null) {
return "Error: SkillManager context not initialized — cannot perform ContentResolver query."
}
val request = JSONObject(result.removePrefix("BRIDGE_REQUEST:"))
val uri = Uri.parse(request.getString("uri"))
val limit = request.optInt("limit", 10)
val columns = arrayOf("_id", "address", "body", "date", "type", "read")
val smsArray = JSONArray()
val cursor = ctx.contentResolver.query(
uri, columns, null, null, "date DESC"
)
cursor?.use {
var count = 0
while (it.moveToNext() && count < limit) {
val row = JSONObject()
row.put("address", it.getString(it.getColumnIndexOrThrow("address")) ?: "")
row.put("body", it.getString(it.getColumnIndexOrThrow("body")) ?: "")
row.put("date", it.getString(it.getColumnIndexOrThrow("date")) ?: "")
row.put("type", it.getString(it.getColumnIndexOrThrow("type")) ?: "1")
row.put("read", it.getString(it.getColumnIndexOrThrow("read")) ?: "0")
smsArray.put(row)
count++
}
}
Log.i("AliceSkills", "SMS bridge: fetched ${smsArray.length()} messages from $uri")
// Re-inject the data and call Python a second time
val injectedArgs = JSONObject(argumentsJson.ifBlank { "{}" })
injectedArgs.put("sms_data", smsArray)
result = executeFunc.call(injectedArgs.toString()).toString()
}
// *** END ADDED
result
} catch (e: Exception) {
Log.e("AliceSkills", "Execution failed for $toolName", e)
"Error executing skill: ${e.message}"
}
}
}
}