mirror of
https://github.com/Aider-AI/aider.git
synced 2025-06-08 05:34:59 +00:00
Try for more cache-friendly prompt ordering
This commit is contained in:
Amar Sood (tekacs)
parent
930880151e
commit
02c092afff
1 changed files with 158 additions and 55 deletions
|
|
@ -15,7 +15,7 @@ from xml.etree.ElementTree import ParseError
|
|||
# Add necessary imports if not already present
|
||||
from collections import defaultdict
|
||||
|
||||
from .base_coder import Coder
|
||||
from .base_coder import Coder, ChatChunks
|
||||
from .editblock_coder import find_original_update_blocks, do_replace, find_similar_lines
|
||||
from .navigator_prompts import NavigatorPrompts
|
||||
from .navigator_legacy_prompts import NavigatorLegacyPrompts
|
||||
|
|
@ -183,63 +183,164 @@ class NavigatorCoder(Coder):
|
|||
Override parent's format_chat_chunks to include enhanced context blocks with a
|
||||
cleaner, more hierarchical structure for better organization.
|
||||
|
||||
Optimized for prompt caching: enhanced context blocks are inserted after static
|
||||
chat elements (system, examples, repo, readonly_files, done) but before variable
|
||||
elements (chat_files, cur, reminder) to preserve prefix caching while providing
|
||||
fresh context information.
|
||||
Optimized for prompt caching by placing context blocks strategically:
|
||||
1. Relatively static blocks (directory structure, environment info) before done_messages
|
||||
2. Dynamic blocks (context summary, symbol outline, git status) after chat_files
|
||||
|
||||
This approach preserves prefix caching while providing fresh context information.
|
||||
"""
|
||||
# First get the normal chat chunks from the parent method
|
||||
chunks = super().format_chat_chunks() # Calls BaseCoder's format_chat_chunks
|
||||
# First get the normal chat chunks from the parent method without calling super
|
||||
# We'll manually build the chunks to control placement of context blocks
|
||||
chunks = self.format_chat_chunks_base()
|
||||
|
||||
# If enhanced context blocks are not enabled, just return the base chunks
|
||||
if not self.use_enhanced_context:
|
||||
return chunks
|
||||
|
||||
# Generate all context blocks
|
||||
env_context = self.get_environment_info()
|
||||
context_summary = self.get_context_summary()
|
||||
dir_structure = self.get_directory_structure()
|
||||
git_status = self.get_git_status()
|
||||
symbol_outline = self.get_context_symbol_outline()
|
||||
|
||||
# 1. Add relatively static blocks BEFORE done_messages
|
||||
# These blocks change less frequently and can be part of the cacheable prefix
|
||||
static_blocks = []
|
||||
if dir_structure:
|
||||
static_blocks.append(dir_structure)
|
||||
if env_context:
|
||||
static_blocks.append(env_context)
|
||||
|
||||
if static_blocks:
|
||||
static_message = "\n\n".join(static_blocks)
|
||||
# Insert as a system message right before done_messages
|
||||
chunks.done.insert(0, dict(role="system", content=static_message))
|
||||
|
||||
# 2. Add dynamic blocks AFTER chat_files
|
||||
# These blocks change with the current files in context
|
||||
dynamic_blocks = []
|
||||
if context_summary:
|
||||
dynamic_blocks.append(context_summary)
|
||||
if symbol_outline:
|
||||
dynamic_blocks.append(symbol_outline)
|
||||
if git_status:
|
||||
dynamic_blocks.append(git_status)
|
||||
|
||||
if dynamic_blocks:
|
||||
dynamic_message = "\n\n".join(dynamic_blocks)
|
||||
# Append as a system message after chat_files
|
||||
chunks.chat_files.append(dict(role="system", content=dynamic_message))
|
||||
|
||||
return chunks
|
||||
|
||||
def format_chat_chunks_base(self):
|
||||
"""
|
||||
Create base chat chunks without enhanced context blocks.
|
||||
This is a copy of the parent's format_chat_chunks method to avoid
|
||||
calling super() which would create a recursive loop.
|
||||
"""
|
||||
self.choose_fence()
|
||||
main_sys = self.fmt_system_prompt(self.gpt_prompts.main_system)
|
||||
|
||||
# If enhanced context blocks are enabled, insert them in a strategic position
|
||||
if self.use_enhanced_context:
|
||||
# Create environment info context block
|
||||
env_context = self.get_environment_info()
|
||||
example_messages = []
|
||||
if self.main_model.examples_as_sys_msg:
|
||||
if self.gpt_prompts.example_messages:
|
||||
main_sys += "\n# Example conversations:\n\n"
|
||||
for msg in self.gpt_prompts.example_messages:
|
||||
role = msg["role"]
|
||||
content = self.fmt_system_prompt(msg["content"])
|
||||
main_sys += f"## {role.upper()}: {content}\n\n"
|
||||
main_sys = main_sys.strip()
|
||||
else:
|
||||
for msg in self.gpt_prompts.example_messages:
|
||||
example_messages.append(
|
||||
dict(
|
||||
role=msg["role"],
|
||||
content=self.fmt_system_prompt(msg["content"]),
|
||||
)
|
||||
)
|
||||
if self.gpt_prompts.example_messages:
|
||||
example_messages += [
|
||||
dict(
|
||||
role="user",
|
||||
content=(
|
||||
"I switched to a new code base. Please don't consider the above files"
|
||||
" or try to edit them any longer."
|
||||
),
|
||||
),
|
||||
dict(role="assistant", content="Ok."),
|
||||
]
|
||||
|
||||
# Get current context summary
|
||||
context_summary = self.get_context_summary()
|
||||
if self.gpt_prompts.system_reminder:
|
||||
main_sys += "\n" + self.fmt_system_prompt(self.gpt_prompts.system_reminder)
|
||||
|
||||
# Get directory structure
|
||||
dir_structure = self.get_directory_structure()
|
||||
chunks = ChatChunks()
|
||||
|
||||
# Get git status
|
||||
git_status = self.get_git_status()
|
||||
if self.main_model.use_system_prompt:
|
||||
chunks.system = [
|
||||
dict(role="system", content=main_sys),
|
||||
]
|
||||
else:
|
||||
chunks.system = [
|
||||
dict(role="user", content=main_sys),
|
||||
dict(role="assistant", content="Ok."),
|
||||
]
|
||||
|
||||
# Get symbol outline for current context files
|
||||
symbol_outline = self.get_context_symbol_outline()
|
||||
chunks.examples = example_messages
|
||||
|
||||
# Collect all context blocks that exist
|
||||
context_blocks = []
|
||||
if env_context:
|
||||
context_blocks.append(env_context)
|
||||
if context_summary:
|
||||
context_blocks.append(context_summary)
|
||||
if dir_structure:
|
||||
context_blocks.append(dir_structure)
|
||||
if git_status:
|
||||
context_blocks.append(git_status)
|
||||
if symbol_outline: # Add the new block if it was generated
|
||||
context_blocks.append(symbol_outline)
|
||||
self.summarize_end()
|
||||
chunks.done = self.done_messages
|
||||
|
||||
# Insert a fresh context update as a separate message before current messages
|
||||
# This preserves cacheable prefix portions (system, examples, repo, etc.)
|
||||
# while still providing fresh context information
|
||||
if context_blocks and chunks.cur:
|
||||
context_message = "\n\n".join(context_blocks)
|
||||
# Insert fresh context as a system message right before the first user message in cur
|
||||
for i, msg in enumerate(chunks.cur):
|
||||
if msg["role"] == "user":
|
||||
# Insert context message right before the first user message
|
||||
chunks.cur.insert(i, dict(role="system", content=context_message))
|
||||
break
|
||||
else:
|
||||
# If no user message found, append to the end of chat_files
|
||||
# (just before any existing cur messages)
|
||||
chunks.chat_files.append(dict(role="system", content=context_message))
|
||||
elif context_blocks:
|
||||
# If there are context blocks but no cur messages, append to chat_files
|
||||
context_message = "\n\n".join(context_blocks)
|
||||
chunks.chat_files.append(dict(role="system", content=context_message))
|
||||
chunks.repo = self.get_repo_messages()
|
||||
chunks.readonly_files = self.get_readonly_files_messages()
|
||||
chunks.chat_files = self.get_chat_files_messages()
|
||||
|
||||
if self.gpt_prompts.system_reminder:
|
||||
reminder_message = [
|
||||
dict(
|
||||
role="system", content=self.fmt_system_prompt(self.gpt_prompts.system_reminder)
|
||||
),
|
||||
]
|
||||
else:
|
||||
reminder_message = []
|
||||
|
||||
chunks.cur = list(self.cur_messages)
|
||||
chunks.reminder = []
|
||||
|
||||
# TODO review impact of token count on image messages
|
||||
messages_tokens = self.main_model.token_count(chunks.all_messages())
|
||||
reminder_tokens = self.main_model.token_count(reminder_message)
|
||||
cur_tokens = self.main_model.token_count(chunks.cur)
|
||||
|
||||
if None not in (messages_tokens, reminder_tokens, cur_tokens):
|
||||
total_tokens = messages_tokens + reminder_tokens + cur_tokens
|
||||
else:
|
||||
# add the reminder anyway
|
||||
total_tokens = 0
|
||||
|
||||
if chunks.cur:
|
||||
final = chunks.cur[-1]
|
||||
else:
|
||||
final = None
|
||||
|
||||
max_input_tokens = self.main_model.info.get("max_input_tokens") or 0
|
||||
# Add the reminder prompt if we still have room to include it.
|
||||
if (
|
||||
not max_input_tokens
|
||||
or total_tokens < max_input_tokens
|
||||
and self.gpt_prompts.system_reminder
|
||||
):
|
||||
if self.main_model.reminder == "sys":
|
||||
chunks.reminder = reminder_message
|
||||
elif self.main_model.reminder == "user" and final and final["role"] == "user":
|
||||
# stuff it into the user message
|
||||
new_content = (
|
||||
final["content"]
|
||||
+ "\n\n"
|
||||
+ self.fmt_system_prompt(self.gpt_prompts.system_reminder)
|
||||
)
|
||||
chunks.cur[-1] = dict(role=final["role"], content=new_content)
|
||||
|
||||
return chunks
|
||||
|
||||
|
|
@ -865,14 +966,16 @@ class NavigatorCoder(Coder):
|
|||
content = params.get('content')
|
||||
after_pattern = params.get('after_pattern')
|
||||
before_pattern = params.get('before_pattern')
|
||||
near_context = params.get('near_context') # New
|
||||
occurrence = params.get('occurrence', 1) # New, default 1
|
||||
occurrence = params.get('occurrence', 1) # Default 1
|
||||
change_id = params.get('change_id')
|
||||
dry_run = params.get('dry_run', False) # New, default False
|
||||
dry_run = params.get('dry_run', False) # Default False
|
||||
position = params.get('position')
|
||||
auto_indent = params.get('auto_indent', True) # Default True
|
||||
use_regex = params.get('use_regex', False) # Default False
|
||||
|
||||
if file_path is not None and content is not None and (after_pattern is not None or before_pattern is not None):
|
||||
if file_path is not None and content is not None and (after_pattern is not None or before_pattern is not None or position is not None):
|
||||
result_message = _execute_insert_block(
|
||||
self, file_path, content, after_pattern, before_pattern, near_context, occurrence, change_id, dry_run
|
||||
self, file_path, content, after_pattern, before_pattern, occurrence, change_id, dry_run, position, auto_indent, use_regex
|
||||
)
|
||||
else:
|
||||
result_message = "Error: Missing required parameters for InsertBlock (file_path, content, and either after_pattern or before_pattern)"
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue