TokioAI-Coder IoT Package v1.0 -- Engine + IoT System Prompt + Tools

4b97c3b verified about 1 month ago

13.7 kB

"""
TokioAI Engine -- Adaptive Memory / Few-Shot Learning
Maintains a persistent store of successful interactions for dynamic few-shot injection.
This is the "fine-tuning in real time" -- instead of modifying weights,
we modify the context the model sees, injecting relevant examples that
worked before.
Storage: JSON file, indexed by intent + keywords for fast retrieval.
"""

import json
import os
import re
import time
from collections import Counter
from typing import Dict, List, Optional

DEFAULT_MEMORY_PATH = os.path.expanduser("~/.tokioai-engine/memory.json")
MAX_EXAMPLES_PER_CATEGORY = 100
MAX_EXAMPLE_AGE_DAYS = 90


class EngineMemory:
    """
    Persistent few-shot memory that makes the model smarter over time.
    
    Stores successful (user_input -> model_output) pairs, tagged by:
    - intent (tool_call, code, text, decision)
    - keywords (extracted from user input)
    - success (was the output valid/useful?)
    - timestamp
    
    On each query, retrieves the most relevant examples for few-shot injection.
    """

    def __init__(self, path: str = None):
        self.path = path or DEFAULT_MEMORY_PATH
        os.makedirs(os.path.dirname(self.path), exist_ok=True)
        self.examples: List[Dict] = []
        self._load()

    def _load(self):
        if os.path.exists(self.path):
            try:
                with open(self.path, "r") as f:
                    self.examples = json.load(f)
            except (json.JSONDecodeError, IOError):
                self.examples = []
        else:
            self.examples = self._seed_examples()
            self._save()

    def _save(self):
        with open(self.path, "w") as f:
            json.dump(self.examples, f, indent=2, ensure_ascii=False)

    def add(
        self,
        user_input: str,
        model_output: str,
        intent: str,
        success: bool = True,
        tool_name: str = None,
        latency_ms: int = None,
    ):
        """Record a successful interaction as a few-shot example."""
        # NEVER store errors, empty outputs, or hallucinated tool output
        if not model_output or len(model_output.strip()) < 5:
            return
        if "Error calling model" in model_output or "Client Error" in model_output:
            return
        if "Server Error" in model_output or "Traceback" in model_output:
            return
        if not success:
            return
        # Skip duplicate: same user input + same first 100 chars of output
        key = (user_input[:100], model_output[:100])
        for ex in self.examples[-50:]:  # Check last 50 only for speed
            if (ex.get("user", "")[:100], ex.get("assistant", "")[:100]) == key:
                return

        keywords = self._extract_keywords(user_input)

        entry = {
            "user": user_input[:500],
            "assistant": model_output[:1000],
            "intent": intent,
            "keywords": keywords,
            "success": success,
            "tool_name": tool_name,
            "latency_ms": latency_ms,
            "timestamp": time.time(),
            "uses": 0,  # How many times this was used as few-shot
        }

        self.examples.append(entry)

        # Prune old/excess entries
        self._prune()
        self._save()

    def retrieve(
        self,
        user_input: str,
        intent: str = None,
        top_k: int = 3,
    ) -> List[Dict]:
        """
        Retrieve the most relevant few-shot examples for a given query.
        Uses keyword overlap + intent matching + recency for ranking.
        """
        if not self.examples:
            return []

        query_keywords = set(self._extract_keywords(user_input))
        scored = []

        for ex in self.examples:
            if not ex.get("success", True):
                continue  # Skip failed examples

            score = 0.0

            # Keyword overlap (most important)
            ex_keywords = set(ex.get("keywords", []))
            overlap = query_keywords & ex_keywords
            if overlap:
                score += len(overlap) * 3.0

            # Intent match
            if intent and ex.get("intent") == intent:
                score += 2.0

            # Recency bonus (newer = better, decay over 30 days)
            age_days = (time.time() - ex.get("timestamp", 0)) / 86400
            recency = max(0, 1.0 - age_days / 30)
            score += recency

            # Usage penalty (avoid showing the same examples every time)
            score -= ex.get("uses", 0) * 0.5

            if score > 0:
                scored.append((score, ex))

        # Sort by score descending
        scored.sort(key=lambda x: x[0], reverse=True)

        # Mark as used
        results = []
        for score, ex in scored[:top_k]:
            ex["uses"] = ex.get("uses", 0) + 1
            results.append(ex)

        if results:
            self._save()

        return results

    def get_stats(self) -> Dict:
        """Return memory statistics."""
        if not self.examples:
            return {"total": 0}

        intents = Counter(ex.get("intent", "unknown") for ex in self.examples)
        tools = Counter(ex.get("tool_name", "none") for ex in self.examples if ex.get("tool_name"))
        success_rate = sum(1 for ex in self.examples if ex.get("success", True)) / len(self.examples)

        return {
            "total": len(self.examples),
            "by_intent": dict(intents),
            "by_tool": dict(tools),
            "success_rate": round(success_rate, 2),
            "oldest_days": round((time.time() - min(ex.get("timestamp", time.time()) for ex in self.examples)) / 86400, 1),
        }

    def feedback(self, index: int, success: bool):
        """Update success status for an example (for learning from failures)."""
        if 0 <= index < len(self.examples):
            self.examples[index]["success"] = success
            self._save()

    def _extract_keywords(self, text: str) -> List[str]:
        """Extract meaningful keywords from text."""
        # Lowercase, split, remove common words
        stopwords = {
            "the", "a", "an", "is", "are", "was", "were", "be", "been",
            "have", "has", "had", "do", "does", "did", "will", "would",
            "could", "should", "may", "might", "can", "shall", "must",
            "to", "of", "in", "for", "on", "with", "at", "by", "from",
            "it", "its", "this", "that", "these", "those", "my", "your",
            "i", "me", "we", "us", "you", "he", "she", "they", "them",
            "and", "or", "but", "not", "if", "then", "else", "so",
            "what", "how", "why", "when", "where", "which", "who",
            "que", "el", "la", "los", "las", "un", "una", "de", "en",
            "por", "para", "con", "es", "no", "si", "como", "me",
            "please", "help", "want", "need", "tell",
        }

        words = re.findall(r'[a-z0-9_./:-]+', text.lower())
        keywords = [w for w in words if w not in stopwords and len(w) > 1]

        # Also extract paths and commands
        paths = re.findall(r'/[\w./+-]+', text)
        keywords.extend(paths)

        return keywords[:20]  # Cap at 20 keywords

    def _prune(self):
        """Remove old or excess entries."""
        now = time.time()
        max_age = MAX_EXAMPLE_AGE_DAYS * 86400

        # Remove old entries
        self.examples = [
            ex for ex in self.examples
            if now - ex.get("timestamp", now) < max_age
        ]

        # If still too many, keep the most successful/used ones
        if len(self.examples) > MAX_EXAMPLES_PER_CATEGORY * 5:
            # Sort by (success, -age) and keep top N
            self.examples.sort(
                key=lambda ex: (ex.get("success", True), -ex.get("timestamp", 0)),
                reverse=True,
            )
            self.examples = self.examples[:MAX_EXAMPLES_PER_CATEGORY * 5]

    def _seed_examples(self) -> List[Dict]:
        """Seed initial few-shot examples for cold start."""
        now = time.time()
        seeds = [
            # Tool call examples
            {
                "user": "check disk space",
                "assistant": '{"name": "execute_local", "arguments": {"command": "df -h"}}',
                "intent": "tool_call",
                "keywords": ["check", "disk", "space"],
                "tool_name": "execute_local",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "what processes are using the most memory",
                "assistant": '{"name": "execute_local", "arguments": {"command": "ps aux --sort=-%mem | head -20"}}',
                "intent": "tool_call",
                "keywords": ["processes", "memory"],
                "tool_name": "execute_local",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "read the nginx config",
                "assistant": '{"name": "read_file", "arguments": {"path": "/etc/nginx/nginx.conf"}}',
                "intent": "tool_call",
                "keywords": ["read", "nginx", "config"],
                "tool_name": "read_file",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "find all python files with TODO comments",
                "assistant": '{"name": "search_files", "arguments": {"pattern": "TODO", "glob": "*.py"}}',
                "intent": "tool_call",
                "keywords": ["find", "python", "todo"],
                "tool_name": "search_files",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "check system health",
                "assistant": '{"name": "diagnose", "arguments": {"target": "local"}}',
                "intent": "tool_call",
                "keywords": ["check", "system", "health"],
                "tool_name": "diagnose",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "show me the contents of /etc/hostname",
                "assistant": '{"name": "read_file", "arguments": {"path": "/etc/hostname"}}',
                "intent": "tool_call",
                "keywords": ["show", "contents", "/etc/hostname"],
                "tool_name": "read_file",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "list docker containers",
                "assistant": '{"name": "execute_local", "arguments": {"command": "docker ps -a"}}',
                "intent": "tool_call",
                "keywords": ["docker", "containers", "list"],
                "tool_name": "execute_local",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "scan ports on 192.168.1.1",
                "assistant": '{"name": "execute_local", "arguments": {"command": "nmap -sV 192.168.1.1"}}',
                "intent": "tool_call",
                "keywords": ["scan", "ports", "nmap", "192.168.1.1"],
                "tool_name": "execute_local",
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            # Text examples (NO tool calls)
            {
                "user": "what is SQL injection",
                "assistant": "SQL injection is an attack where malicious SQL code is inserted into application queries through user input. The attacker can read, modify, or delete database data, bypass authentication, or even execute system commands. Prevention: parameterized queries, input validation, least-privilege DB accounts, WAF rules.",
                "intent": "text",
                "keywords": ["sql", "injection"],
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "explain the TCP 3-way handshake",
                "assistant": "TCP 3-way handshake establishes a connection: (1) Client sends SYN with initial sequence number, (2) Server responds SYN-ACK with its own sequence number and acknowledges client's, (3) Client sends ACK. After this, data flows bidirectionally. Termination uses a 4-way FIN handshake.",
                "intent": "text",
                "keywords": ["tcp", "handshake"],
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            {
                "user": "hi",
                "assistant": "Hey. What do you need?",
                "intent": "text",
                "keywords": ["hi"],
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
            # Decision examples (model correctly chose NOT to use tool)
            {
                "user": "what is docker",
                "assistant": "Docker is a containerization platform that packages applications with their dependencies into isolated containers. Containers share the host kernel (unlike VMs), making them lightweight and fast. Key concepts: images (blueprints), containers (running instances), Dockerfile (build instructions), docker-compose (multi-container apps).",
                "intent": "decision",
                "keywords": ["docker"],
                "success": True,
                "timestamp": now,
                "uses": 0,
            },
        ]
        return seeds

	"""
	TokioAI Engine -- Adaptive Memory / Few-Shot Learning
	Maintains a persistent store of successful interactions for dynamic few-shot injection.

	This is the "fine-tuning in real time" -- instead of modifying weights,
	we modify the context the model sees, injecting relevant examples that
	worked before.

	Storage: JSON file, indexed by intent + keywords for fast retrieval.
	"""

	import json
	import os
	import re
	import time
	from collections import Counter
	from typing import Dict, List, Optional

	DEFAULT_MEMORY_PATH = os.path.expanduser("~/.tokioai-engine/memory.json")
	MAX_EXAMPLES_PER_CATEGORY = 100
	MAX_EXAMPLE_AGE_DAYS = 90


	class EngineMemory:
	"""
	Persistent few-shot memory that makes the model smarter over time.

	Stores successful (user_input -> model_output) pairs, tagged by:
	- intent (tool_call, code, text, decision)
	- keywords (extracted from user input)
	- success (was the output valid/useful?)
	- timestamp

	On each query, retrieves the most relevant examples for few-shot injection.
	"""

	def __init__(self, path: str = None):
	self.path = path or DEFAULT_MEMORY_PATH
	os.makedirs(os.path.dirname(self.path), exist_ok=True)
	self.examples: List[Dict] = []
	self._load()

	def _load(self):
	if os.path.exists(self.path):
	try:
	with open(self.path, "r") as f:
	self.examples = json.load(f)
	except (json.JSONDecodeError, IOError):
	self.examples = []
	else:
	self.examples = self._seed_examples()
	self._save()

	def _save(self):
	with open(self.path, "w") as f:
	json.dump(self.examples, f, indent=2, ensure_ascii=False)

	def add(
	self,
	user_input: str,
	model_output: str,
	intent: str,
	success: bool = True,
	tool_name: str = None,
	latency_ms: int = None,
	):
	"""Record a successful interaction as a few-shot example."""
	# NEVER store errors, empty outputs, or hallucinated tool output
	if not model_output or len(model_output.strip()) < 5:
	return
	if "Error calling model" in model_output or "Client Error" in model_output:
	return
	if "Server Error" in model_output or "Traceback" in model_output:
	return
	if not success:
	return
	# Skip duplicate: same user input + same first 100 chars of output
	key = (user_input[:100], model_output[:100])
	for ex in self.examples[-50:]: # Check last 50 only for speed
	if (ex.get("user", "")[:100], ex.get("assistant", "")[:100]) == key:
	return

	keywords = self._extract_keywords(user_input)

	entry = {
	"user": user_input[:500],
	"assistant": model_output[:1000],
	"intent": intent,
	"keywords": keywords,
	"success": success,
	"tool_name": tool_name,
	"latency_ms": latency_ms,
	"timestamp": time.time(),
	"uses": 0, # How many times this was used as few-shot
	}

	self.examples.append(entry)

	# Prune old/excess entries
	self._prune()
	self._save()

	def retrieve(
	self,
	user_input: str,
	intent: str = None,
	top_k: int = 3,
	) -> List[Dict]:
	"""
	Retrieve the most relevant few-shot examples for a given query.
	Uses keyword overlap + intent matching + recency for ranking.
	"""
	if not self.examples:
	return []

	query_keywords = set(self._extract_keywords(user_input))
	scored = []

	for ex in self.examples:
	if not ex.get("success", True):
	continue # Skip failed examples

	score = 0.0

	# Keyword overlap (most important)
	ex_keywords = set(ex.get("keywords", []))
	overlap = query_keywords & ex_keywords
	if overlap:
	score += len(overlap) * 3.0

	# Intent match
	if intent and ex.get("intent") == intent:
	score += 2.0

	# Recency bonus (newer = better, decay over 30 days)
	age_days = (time.time() - ex.get("timestamp", 0)) / 86400
	recency = max(0, 1.0 - age_days / 30)
	score += recency

	# Usage penalty (avoid showing the same examples every time)
	score -= ex.get("uses", 0) * 0.5

	if score > 0:
	scored.append((score, ex))

	# Sort by score descending
	scored.sort(key=lambda x: x[0], reverse=True)

	# Mark as used
	results = []
	for score, ex in scored[:top_k]:
	ex["uses"] = ex.get("uses", 0) + 1
	results.append(ex)

	if results:
	self._save()

	return results

	def get_stats(self) -> Dict:
	"""Return memory statistics."""
	if not self.examples:
	return {"total": 0}

	intents = Counter(ex.get("intent", "unknown") for ex in self.examples)
	tools = Counter(ex.get("tool_name", "none") for ex in self.examples if ex.get("tool_name"))
	success_rate = sum(1 for ex in self.examples if ex.get("success", True)) / len(self.examples)

	return {
	"total": len(self.examples),
	"by_intent": dict(intents),
	"by_tool": dict(tools),
	"success_rate": round(success_rate, 2),
	"oldest_days": round((time.time() - min(ex.get("timestamp", time.time()) for ex in self.examples)) / 86400, 1),
	}

	def feedback(self, index: int, success: bool):
	"""Update success status for an example (for learning from failures)."""
	if 0 <= index < len(self.examples):
	self.examples[index]["success"] = success
	self._save()

	def _extract_keywords(self, text: str) -> List[str]:
	"""Extract meaningful keywords from text."""
	# Lowercase, split, remove common words
	stopwords = {
	"the", "a", "an", "is", "are", "was", "were", "be", "been",
	"have", "has", "had", "do", "does", "did", "will", "would",
	"could", "should", "may", "might", "can", "shall", "must",
	"to", "of", "in", "for", "on", "with", "at", "by", "from",
	"it", "its", "this", "that", "these", "those", "my", "your",
	"i", "me", "we", "us", "you", "he", "she", "they", "them",
	"and", "or", "but", "not", "if", "then", "else", "so",
	"what", "how", "why", "when", "where", "which", "who",
	"que", "el", "la", "los", "las", "un", "una", "de", "en",
	"por", "para", "con", "es", "no", "si", "como", "me",
	"please", "help", "want", "need", "tell",
	}

	words = re.findall(r'[a-z0-9_./:-]+', text.lower())
	keywords = [w for w in words if w not in stopwords and len(w) > 1]

	# Also extract paths and commands
	paths = re.findall(r'/[\w./+-]+', text)
	keywords.extend(paths)

	return keywords[:20] # Cap at 20 keywords

	def _prune(self):
	"""Remove old or excess entries."""
	now = time.time()
	max_age = MAX_EXAMPLE_AGE_DAYS * 86400

	# Remove old entries
	self.examples = [
	ex for ex in self.examples
	if now - ex.get("timestamp", now) < max_age
	]

	# If still too many, keep the most successful/used ones
	if len(self.examples) > MAX_EXAMPLES_PER_CATEGORY * 5:
	# Sort by (success, -age) and keep top N
	self.examples.sort(
	key=lambda ex: (ex.get("success", True), -ex.get("timestamp", 0)),
	reverse=True,
	)
	self.examples = self.examples[:MAX_EXAMPLES_PER_CATEGORY * 5]

	def _seed_examples(self) -> List[Dict]:
	"""Seed initial few-shot examples for cold start."""
	now = time.time()
	seeds = [
	# Tool call examples
	{
	"user": "check disk space",
	"assistant": '{"name": "execute_local", "arguments": {"command": "df -h"}}',
	"intent": "tool_call",
	"keywords": ["check", "disk", "space"],
	"tool_name": "execute_local",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "what processes are using the most memory",
	"assistant": '{"name": "execute_local", "arguments": {"command": "ps aux --sort=-%mem \| head -20"}}',
	"intent": "tool_call",
	"keywords": ["processes", "memory"],
	"tool_name": "execute_local",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "read the nginx config",
	"assistant": '{"name": "read_file", "arguments": {"path": "/etc/nginx/nginx.conf"}}',
	"intent": "tool_call",
	"keywords": ["read", "nginx", "config"],
	"tool_name": "read_file",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "find all python files with TODO comments",
	"assistant": '{"name": "search_files", "arguments": {"pattern": "TODO", "glob": "*.py"}}',
	"intent": "tool_call",
	"keywords": ["find", "python", "todo"],
	"tool_name": "search_files",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "check system health",
	"assistant": '{"name": "diagnose", "arguments": {"target": "local"}}',
	"intent": "tool_call",
	"keywords": ["check", "system", "health"],
	"tool_name": "diagnose",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "show me the contents of /etc/hostname",
	"assistant": '{"name": "read_file", "arguments": {"path": "/etc/hostname"}}',
	"intent": "tool_call",
	"keywords": ["show", "contents", "/etc/hostname"],
	"tool_name": "read_file",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "list docker containers",
	"assistant": '{"name": "execute_local", "arguments": {"command": "docker ps -a"}}',
	"intent": "tool_call",
	"keywords": ["docker", "containers", "list"],
	"tool_name": "execute_local",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "scan ports on 192.168.1.1",
	"assistant": '{"name": "execute_local", "arguments": {"command": "nmap -sV 192.168.1.1"}}',
	"intent": "tool_call",
	"keywords": ["scan", "ports", "nmap", "192.168.1.1"],
	"tool_name": "execute_local",
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	# Text examples (NO tool calls)
	{
	"user": "what is SQL injection",
	"assistant": "SQL injection is an attack where malicious SQL code is inserted into application queries through user input. The attacker can read, modify, or delete database data, bypass authentication, or even execute system commands. Prevention: parameterized queries, input validation, least-privilege DB accounts, WAF rules.",
	"intent": "text",
	"keywords": ["sql", "injection"],
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "explain the TCP 3-way handshake",
	"assistant": "TCP 3-way handshake establishes a connection: (1) Client sends SYN with initial sequence number, (2) Server responds SYN-ACK with its own sequence number and acknowledges client's, (3) Client sends ACK. After this, data flows bidirectionally. Termination uses a 4-way FIN handshake.",
	"intent": "text",
	"keywords": ["tcp", "handshake"],
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	{
	"user": "hi",
	"assistant": "Hey. What do you need?",
	"intent": "text",
	"keywords": ["hi"],
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	# Decision examples (model correctly chose NOT to use tool)
	{
	"user": "what is docker",
	"assistant": "Docker is a containerization platform that packages applications with their dependencies into isolated containers. Containers share the host kernel (unlike VMs), making them lightweight and fast. Key concepts: images (blueprints), containers (running instances), Dockerfile (build instructions), docker-compose (multi-container apps).",
	"intent": "decision",
	"keywords": ["docker"],
	"success": True,
	"timestamp": now,
	"uses": 0,
	},
	]
	return seeds