| |
| |
| @@ -1,3 +1,4 @@ |
| +# Modified. |
| import json |
| import time |
| import traceback |
| @@ -70,6 +71,33 @@ |
| logger = get_logger(__name__) |
| |
| |
| +def _fit_context_window(sequence: List[Message], new_messages: List[Message], limit: Optional[int]) -> List[Message]: |
| + if not limit or len(sequence) <= 2 or not new_messages: |
| + return sequence |
| + counts = get_token_counts_for_messages(sequence) |
| + if sum(counts) <= limit: |
| + return sequence |
| + n_new = len(new_messages) |
| + room = limit - counts[0] - sum(counts[-n_new:]) |
| + middle = list(zip(sequence[1:-n_new], counts[1:-n_new])) |
| + units, i = [], 0 |
| + while i < len(middle): |
| + j = i + 1 |
| + if getattr(middle[i][0], "role", None) == "assistant" and middle[i][0].tool_calls: |
| + while j < len(middle) and getattr(middle[j][0], "role", None) == "tool": |
| + j += 1 |
| + units.append(middle[i:j]) |
| + i = j |
| + kept = [] |
| + for unit in reversed(units): |
| + size = sum(n for _, n in unit) |
| + if size > room: |
| + break |
| + kept[:0] = [m for m, _ in unit] |
| + room -= size |
| + return [sequence[0], *kept, *new_messages] |
| + |
| + |
| class BaseAgent(ABC): |
| """ |
| Abstract class for all agents. |
| @@ -745,6 +773,23 @@ |
| elif max_chaining_steps is not None and counter > max_chaining_steps: |
| self.logger.info(f"Hit max chaining steps, stopping after {counter} steps") |
| break |
| + elif (max_chaining_steps is not None and counter == max_chaining_steps |
| + and (heartbeat_request or function_failed or token_warning) |
| + and __import__("os").environ.get("MEMEBENCH_LETTA_FINAL_REPLY_AT_CAP") == "1"): |
| + assert self.agent_state.created_by_id is not None |
| + next_input_messages = [ |
| + Message.dict_to_message( |
| + agent_id=self.agent_state.id, |
| + model=self.model, |
| + openai_message_dict={ |
| + "role": "user", |
| + "content": get_heartbeat( |
| + "You have reached the maximum number of steps. Call send_message now " |
| + "with your final response."), |
| + }, |
| + ) |
| + ] |
| + continue |
| # Chain handlers |
| elif token_warning and summarizer_settings.send_memory_warning_message: |
| assert self.agent_state.created_by_id is not None |
| @@ -828,6 +873,8 @@ |
| |
| in_context_messages = self.agent_manager.get_in_context_messages(agent_id=self.agent_state.id, actor=self.user) |
| input_message_sequence = in_context_messages + messages |
| + input_message_sequence = _fit_context_window( |
| + input_message_sequence, messages, self.agent_state.llm_config.context_window) |
| |
| if ( |
| len(input_message_sequence) > 1 |
| |
| |
| @@ -1,7 +1,8 @@ |
| +# Modified. |
| import os |
| from logging import CRITICAL, DEBUG, ERROR, INFO, NOTSET, WARN, WARNING |
| |
| -LETTA_DIR = os.path.join(os.path.expanduser("~"), ".letta") |
| +LETTA_DIR = os.environ.get("LETTA_DIR") or os.path.join(os.path.expanduser("~"), ".letta") |
| LETTA_TOOL_EXECUTION_DIR = os.path.join(LETTA_DIR, "tool_execution_dir") |
| |
| ADMIN_PREFIX = "/v1/admin" |
| |
| |
| @@ -1,3 +1,4 @@ |
| +# Modified. |
| from typing import Optional |
| |
| from letta.agent import Agent |
| @@ -43,7 +44,7 @@ |
| page = int(page) |
| except: |
| raise ValueError(f"'page' argument must be an integer") |
| - count = RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE |
| + count = (self.agent_state.metadata or {}).get("retrieve_num", RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE) |
| # TODO: add paging by page number. currently cursor only works with strings. |
| # original: start=page * count |
| messages = self.message_manager.list_user_messages_for_agent( |
| @@ -104,7 +105,7 @@ |
| page = int(page) |
| except: |
| raise ValueError(f"'page' argument must be an integer") |
| - count = RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE |
| + count = (self.agent_state.metadata or {}).get("retrieve_num", RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE) |
| |
| try: |
| # Get results using passage manager |
| |
| |
| @@ -1,3 +1,4 @@ |
| +# Modified. |
| from typing import Any, Dict, List, Literal, Optional, Union |
| |
| from pydantic import BaseModel, Field, field_validator |
| @@ -124,6 +125,7 @@ |
| logprobs: Optional[bool] = False |
| top_logprobs: Optional[int] = None |
| max_tokens: Optional[int] = None |
| + max_completion_tokens: Optional[int] = None |
| n: Optional[int] = 1 |
| presence_penalty: Optional[float] = 0 |
| response_format: Optional[ResponseFormat] = None |
| |
| |
| @@ -1,7 +1,8 @@ |
| +# Modified. |
| import json |
| from typing import List, Optional, Sequence |
| |
| -from sqlalchemy import delete, exists, func, select, text |
| +from sqlalchemy import String, cast, delete, exists, func, select, text |
| |
| from letta.log import get_logger |
| from letta.orm.agent import Agent as AgentModel |
| @@ -304,7 +305,9 @@ |
| query = query.filter(MessageModel.group_id == group_id) |
| |
| # If query_text is provided, filter messages using subquery + json_array_elements. |
| - if query_text: |
| + if query_text and session.bind is not None and session.bind.dialect.name == "sqlite": |
| + query = query.filter(func.lower(cast(MessageModel.content, String)).like(f"%{query_text.lower()}%")) |
| + elif query_text: |
| content_element = func.json_array_elements(MessageModel.content).alias("content_element") |
| query = query.filter( |
| exists( |
|
|
