File size: 21,714 Bytes
9fced79
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
import os
from typing import List, Dict, Any, Optional, Type, Callable
from datetime import datetime, timedelta
import heapq
import json

import torch
from langchain_core.tools import BaseTool
from langchain_core.language_models import BaseChatModel
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.messages import SystemMessage, HumanMessage, AIMessage, ToolMessage
from langchain_core.vectorstores import VectorStore
from langchain_core.documents import Document
from langchain_core.output_parsers import StrOutputParser
from langchain.tools.retriever import create_retriever_tool
from langchain_huggingface import HuggingFaceEmbeddings

from langgraph.graph import StateGraph, END
from langgraph.prebuilt import (
    ToolNode, 
    ToolInvocation, 
    agent_executor, 
    create_function_calling_executor,
    AgentState,
    MessageGraph
)
from langgraph.prebuilt.tool_executor import ToolExecutor, extract_tool_invocations
from langgraph.prebuilt.tool_nodes import get_default_tool_node_parser


class AdvancedToolAgent:
    """
    An advanced agent with robust tool-calling capabilities using LangGraph.
    Features enhanced memory management, context enrichment, and tool execution tracking.
    """
    
    def __init__(
        self, 
        embedding_model: HuggingFaceEmbeddings, 
        vector_store: VectorStore, 
        llm: BaseChatModel,
        tools: Optional[List[BaseTool]] = None,
        max_iterations: int = 10,
        memory_threshold: float = 0.7
    ):
        """
        Initialize the agent with required components.
        
        Args:
            embedding_model: Model for embedding text
            vector_store: Storage for agent memory
            llm: Language model for agent reasoning
            tools: List of tools accessible to the agent
            max_iterations: Maximum number of tool calling iterations
            memory_threshold: Threshold for deciding when to include memory context (0-1)
        """
        self.embedding_model = embedding_model
        self.vector_store = vector_store
        self.llm = llm
        self.tools = tools or []
        self.max_iterations = max_iterations
        self.memory_threshold = memory_threshold
        
        # Setup retriever for memory access
        self.retriever = vector_store.as_retriever(
            search_kwargs={"k": 3, "score_threshold": 0.75}
        )
        
        # Create memory retrieval tool
        self.memory_tool = create_retriever_tool(
            retriever=self.retriever,
            name="memory_search",
            description="Search the agent's memory for relevant past interactions and knowledge."
        )
        
        # Add memory tool to the agent's toolset
        self.all_tools = self.tools + [self.memory_tool]
        
        # Setup tool executor
        self.tool_executor = ToolExecutor(self.all_tools)
        
        # Build the agent's execution graph
        self.agent_executor = self._build_agent_graph()
        
        print(f"AdvancedToolAgent initialized with {len(self.all_tools)} tools")

    def __call__(self, question: str) -> str:
        """
        Process a question using the agent.
        
        Args:
            question: The user query to respond to
            
        Returns:
            The agent's response
        """
        print(f"Agent received question: {question[:50]}..." if len(question) > 50 else question)
        
        # Enrich context with relevant memory
        enriched_input = self._enrich_context(question)
        
        # Create initial state
        initial_state = {
            "messages": [HumanMessage(content=enriched_input)],
            "tools": self.all_tools,
            "tool_calls": [],
        }
        
        # Execute agent graph
        final_state = self.agent_executor.invoke(initial_state)
        
        # Extract the final response
        final_message = final_state["messages"][-1]
        answer = final_message.content
        
        # Store this interaction in memory
        self._store_interaction(question, answer, final_state.get("tool_calls", []))
        
        # Periodically manage memory
        self._periodic_memory_management()
        
        print(f"Agent returning answer: {answer[:50]}..." if len(answer) > 50 else answer)
        return answer

    def _build_agent_graph(self):
        """Build the LangGraph execution graph with enhanced tool calling"""
        
        # Function for the agent to process messages and call tools
        def agent_node(state: AgentState) -> AgentState:
            """Process messages and decide on next action"""
            messages = state["messages"]
            
            # Add system instructions with tool details
            if not any(isinstance(msg, SystemMessage) for msg in messages):
                system_prompt = self._create_system_prompt()
                messages = [SystemMessage(content=system_prompt)] + messages
            
            # Get response from LLM
            response = self.llm.invoke(messages)
            
            # Extract any tool calls
            tool_calls = extract_tool_invocations(
                response,
                self.all_tools,
                strict_mode=False,
            )
            
            # Update state
            new_state = state.copy()
            new_state["messages"] = messages + [response]
            new_state["tool_calls"] = tool_calls
            
            return new_state
        
        # Function for executing tools
        def tool_node(state: AgentState) -> AgentState:
            """Execute tools and add results to messages"""
            # Get the tool calls from the state
            tool_calls = state["tool_calls"]
            
            # Execute each tool call
            tool_results = []
            for tool_call in tool_calls:
                try:
                    # Execute the tool
                    result = self.tool_executor.invoke(tool_call)
                    
                    # Create a tool message with the result
                    tool_msg = ToolMessage(
                        content=str(result),
                        tool_call_id=tool_call.id,
                        name=tool_call.name,
                    )
                    tool_results.append(tool_msg)
                    
                    # Track tool usage for memory
                    self._track_tool_usage(tool_call.name, tool_call.args, result)
                except Exception as e:
                    # Handle tool execution errors
                    error_msg = f"Error executing tool {tool_call.name}: {str(e)}"
                    tool_msg = ToolMessage(
                        content=error_msg,
                        tool_call_id=tool_call.id,
                        name=tool_call.name,
                    )
                    tool_results.append(tool_msg)
            
            # Update state with tool results
            new_state = state.copy()
            new_state["messages"] = state["messages"] + tool_results
            new_state["tool_calls"] = []
            return new_state
        
        # Create the graph
        graph = StateGraph(AgentState)
        
        # Add nodes
        graph.add_node("agent", agent_node)
        graph.add_node("tools", tool_node)
        
        # Set the entry point
        graph.set_entry_point("agent")
        
        # Add edges
        graph.add_conditional_edges(
            "agent",
            lambda state: "tools" if state["tool_calls"] else END,
            {
                "tools": "tools",
                END: END,
            }
        )
        graph.add_edge("tools", "agent")
        
        # Set max iterations to prevent infinite loops
        return graph.compile(max_iterations=self.max_iterations)

    def _create_system_prompt(self) -> str:
        """Create a system prompt with tool instructions"""
        tool_descriptions = "\n\n".join([
            f"Tool {i+1}: {tool.name}\n"
            f"Description: {tool.description}\n"
            f"Args: {json.dumps(tool.args, indent=2) if hasattr(tool, 'args') else 'No arguments required'}"
            for i, tool in enumerate(self.all_tools)
        ])
        
        return f"""You are an advanced AI assistant with access to various tools.
When a user asks a question, use your knowledge and the available tools to provide
accurate and helpful responses.

AVAILABLE TOOLS:
{tool_descriptions}

INSTRUCTIONS FOR TOOL USAGE:
1. When you need information that requires a tool, call the appropriate tool.
2. Format tool calls clearly by specifying the tool name and inputs.
3. Wait for tool results before providing final answers.
4. Use tools only when necessary - if you can answer directly, do so.
5. If a tool fails, try a different approach or tool.
6. Always explain your reasoning step by step.

Remember to be helpful, accurate, and concise in your responses.
"""

    def _enrich_context(self, query: str) -> str:
        """Enrich the input query with relevant context from memory"""
        # Search for similar content
        similar_docs = self.vector_store.similarity_search(
            query, 
            k=2,  # Limit to 2 most relevant documents
            fetch_k=5  # Consider 5 candidates
        )
        
        # Only use memory if relevance is high enough
        if not similar_docs or len(similar_docs) == 0:
            return query
            
        # Build enhanced context
        context_additions = []
        for doc in similar_docs:
            content = doc.page_content
            
            # Extract different types of memory
            if "Question:" in content and "Final answer:" in content:
                # Q&A memory
                q = content.split("Question:")[1].split("Final answer:")[0].strip()
                a = content.split("Final answer:")[1].split("Timestamp:", 1)[0].strip()
                
                # Only add if it's not too similar to current question
                if not self._is_similar_question(query, q, threshold=0.85):
                    context_additions.append(f"Related Q: {q}\nRelated A: {a}")
                    
            elif "Tool Knowledge" in content:
                # Tool usage memory
                tool_name = content.split("Tool:")[1].split("Query:")[0].strip()
                tool_result = content.split("Result:")[1].split("Timestamp:")[0].strip()
                context_additions.append(
                    f"From prior tool use ({tool_name}): {tool_result[:200]}"
                )
        
        # Only add context if we have relevant information
        if context_additions:
            return (
                "Consider this relevant information first:\n\n" + 
                "\n\n".join(context_additions[:2]) +  # Limit to 2 pieces of context
                "\n\nNow answering this question: " + query
            )
        else:
            return query

    def _is_similar_question(self, query1: str, query2: str, threshold: float = 0.8) -> bool:
        """Check if two questions are semantically similar using embeddings"""
        # Get embeddings for both queries
        if hasattr(self.embedding_model, 'embed_query'):
            emb1 = self.embedding_model.embed_query(query1)
            emb2 = self.embedding_model.embed_query(query2)
            
            # Calculate cosine similarity
            similarity = self._cosine_similarity(emb1, emb2)
            return similarity > threshold
        return False

    @staticmethod
    def _cosine_similarity(v1, v2):
        """Calculate cosine similarity between vectors"""
        dot_product = sum(x * y for x, y in zip(v1, v2))
        magnitude1 = sum(x * x for x in v1) ** 0.5
        magnitude2 = sum(x * x for x in v2) ** 0.5
        if magnitude1 * magnitude2 == 0:
            return 0
        return dot_product / (magnitude1 * magnitude2)

    def _store_interaction(self, question: str, answer: str, tool_calls: List[dict]) -> None:
        """Store the interaction in vector memory"""
        timestamp = datetime.now().isoformat()
        
        # Format tools used
        tools_used = []
        for tool_call in tool_calls:
            if isinstance(tool_call, dict) and 'name' in tool_call:
                tools_used.append(tool_call['name'])
            elif hasattr(tool_call, 'name'):
                tools_used.append(tool_call.name)
                
        tools_str = ", ".join(tools_used) if tools_used else "None"
        
        # Create content
        content = (
            f"Question: {question}\n"
            f"Tools Used: {tools_str}\n"
            f"Final answer: {answer}\n"
            f"Timestamp: {timestamp}"
        )
        
        # Create document with metadata
        doc = Document(
            page_content=content,
            metadata={
                "question": question,
                "timestamp": timestamp,
                "type": "qa_pair",
                "tools_used": tools_str
            }
        )
        
        # Add to vector store
        self.vector_store.add_documents([doc])

    def _track_tool_usage(self, tool_name: str, tool_input: Any, tool_output: Any) -> None:
        """Track tool usage for future reference"""
        timestamp = datetime.now().isoformat()
        
        # Format the content
        content = (
            f"Tool Knowledge\n"
            f"Tool: {tool_name}\n"
            f"Query: {str(tool_input)}\n"
            f"Result: {str(tool_output)}\n"
            f"Timestamp: {timestamp}"
        )
        
        # Create document with metadata
        doc = Document(
            page_content=content,
            metadata={
                "type": "tool_knowledge",
                "tool": tool_name,
                "timestamp": timestamp
            }
        )
        
        # Add to vector store
        self.vector_store.add_documents([doc])

    def _periodic_memory_management(self, 
                                   check_frequency: int = 10, 
                                   max_documents: int = 1000, 
                                   max_age_days: int = 30) -> None:
        """Periodically manage memory to prevent unbounded growth"""
        # Simple probabilistic check to avoid running this too often
        if hash(datetime.now().isoformat()) % check_frequency != 0:
            return
            
        self.manage_memory(max_documents, max_age_days)

    def manage_memory(self, max_documents: int = 1000, max_age_days: int = 30) -> None:
        """
        Manage memory by pruning old or less useful entries from the vector store.
        
        Args:
            max_documents: Maximum number of documents to keep
            max_age_days: Remove documents older than this many days
        """
        print(f"Starting memory management...")
        
        # Get all documents from the vector store
        try:
            # For vector stores that have a get_all_documents method
            if hasattr(self.vector_store, "get_all_documents"):
                all_docs = self.vector_store.get_all_documents()
                all_ids = [doc.metadata.get("id", i) for i, doc in enumerate(all_docs)]
            # For other vector store implementations
            else:
                print("Warning: Vector store doesn't expose required attributes for memory management")
                return
        except Exception as e:
            print(f"Error accessing vector store documents: {e}")
            return
        
        if not all_docs:
            print("No documents found in vector store")
            return
            
        print(f"Retrieved {len(all_docs)} documents for scoring")
        
        # Score each document based on recency, importance and relevance
        scored_docs = []
        cutoff_date = datetime.now() - timedelta(days=max_age_days)
        
        for i, doc in enumerate(all_docs):
            doc_id = all_ids[i] if i < len(all_ids) else i
            
            # Extract timestamp from content or metadata
            timestamp = None
            if hasattr(doc, "metadata") and doc.metadata and "timestamp" in doc.metadata:
                try:
                    timestamp = datetime.fromisoformat(doc.metadata["timestamp"])
                except (ValueError, TypeError):
                    pass
            
            # If no timestamp in metadata, try to extract from content
            if not timestamp and hasattr(doc, "page_content") and "Timestamp:" in doc.page_content:
                try:
                    timestamp_str = doc.page_content.split("Timestamp:")[-1].strip().split('\n')[0]
                    timestamp = datetime.fromisoformat(timestamp_str)
                except (ValueError, TypeError):
                    timestamp = datetime.now() - timedelta(days=max_age_days+1)
            
            # If still no timestamp, use a default
            if not timestamp:
                timestamp = datetime.now() - timedelta(days=max_age_days+1)
            
            # Calculate age score (newer is better)
            age_factor = max(0.0, min(1.0, (timestamp - cutoff_date).total_seconds() / 
                                     (datetime.now() - cutoff_date).total_seconds()))
            
            # Calculate importance score based on document type and access frequency
            importance_factor = 1.0
            
            # Tool knowledge is more valuable
            if hasattr(doc, "metadata") and doc.metadata and doc.metadata.get("type") == "tool_knowledge":
                importance_factor += 0.5
                
            # If document has been accessed often, increase importance
            if hasattr(doc, "metadata") and doc.metadata and "access_count" in doc.metadata:
                importance_factor += min(1.0, doc.metadata["access_count"] / 10)
                
            # If document contains references to complex tools, prioritize it
            if hasattr(doc, "page_content"):
                complex_tools = ["web_search", "python_repl", "analyze_image", "arxiv_search"]
                if any(tool in doc.page_content for tool in complex_tools):
                    importance_factor += 0.3
            
            # Create combined score (higher = more valuable to keep)
            total_score = (0.6 * age_factor) + (0.4 * importance_factor)
            
            # Add to priority queue (negative for max-heap behavior)
            heapq.heappush(scored_docs, (-total_score, i, doc))
        
        # Select top documents to keep
        docs_to_keep = []
        for _ in range(min(max_documents, len(scored_docs))):
            if scored_docs:
                _, _, doc = heapq.heappop(scored_docs)
                docs_to_keep.append(doc)
        
        # Only rebuild if we're actually pruning some documents
        if len(docs_to_keep) < len(all_docs):
            print(f"Memory management: Keeping {len(docs_to_keep)} documents out of {len(all_docs)}")
            
            # Create a new vector store with the same type as the current one
            vector_store_type = type(self.vector_store)
            
            # Different approaches based on vector store type
            if hasattr(vector_store_type, "from_documents"):
                # Most langchain vector stores support this method
                new_vector_store = vector_store_type.from_documents(
                    docs_to_keep,
                    embedding=self.embedding_model
                )
                self.vector_store = new_vector_store
                print(f"Vector store rebuilt with {len(docs_to_keep)} documents")
                
            elif hasattr(vector_store_type, "from_texts"):
                # For vector stores that use from_texts
                texts = [doc.page_content for doc in docs_to_keep]
                metadatas = [doc.metadata if hasattr(doc, "metadata") else {} for doc in docs_to_keep]
                
                new_vector_store = vector_store_type.from_texts(
                    texts=texts,
                    embedding=self.embedding_model,
                    metadatas=metadatas
                )
                self.vector_store = new_vector_store
                print(f"Vector store rebuilt with {len(docs_to_keep)} documents")
                
            else:
                print("Warning: Could not determine how to rebuild the vector store")
                print(f"Vector store type: {vector_store_type.__name__}")

# Example usage
if __name__ == "__main__":
    from langchain_huggingface import HuggingFaceEmbeddings
    from langchain_chroma import Chroma
    from langchain_groq import ChatGroq
    from basic_tools import multiply, add, subtract, divide, wiki_search, web_search
    
    # Initialize embeddings
    embeddings = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-mpnet-base-v2",
        model_kwargs={"device": "cuda" if torch.cuda.is_available() else "cpu"}
    )
    
    # Initialize vector store
    vector_store = Chroma(
        embedding_function=embeddings,
        collection_name="advanced_agent_memory"
    )
    
    # Initialize LLM
    llm = ChatGroq(model="qwen-qwq-32b", temperature=0)
    
    # Define tools
    tools = [multiply, add, subtract, divide, wiki_search, web_search]
    
    # Create agent
    agent = AdvancedToolAgent(
        embedding_model=embeddings,
        vector_store=vector_store,
        llm=llm,
        tools=tools
    )
    
    # Test the agent
    response = agent("What is the population of France multiplied by 2?")
    print(f"Response: {response}")