Merge pull request #437 from Codium-ai/tr/new_gpt4

Introduce support for 'gpt-4-1106-preview' model and dynamic token limit calculation

pr_agent/algo/__init__.py

@@ -8,6 +8,7 @@ MAX_TOKENS = {
     'gpt-4': 8000,
     'gpt-4-0613': 8000,
     'gpt-4-32k': 32000,
+    'gpt-4-1106-preview': 128000, # 128K, but may be limited by config.max_model_tokens
     'claude-instant-1': 100000,
     'claude-2': 100000,
     'command-nightly': 4096,

pr_agent/algo/pr_processing.py

@@ -7,11 +7,11 @@ from typing import Any, Callable, List, Tuple
 
 from github import RateLimitExceededException
 
-from pr_agent.algo import MAX_TOKENS
 from pr_agent.algo.git_patch_processing import convert_to_hunks_with_lines_numbers, extend_patch, handle_patch_deletions
 from pr_agent.algo.language_handler import sort_files_by_main_languages
 from pr_agent.algo.file_filter import filter_ignored
 from pr_agent.algo.token_handler import TokenHandler, get_token_encoder
+from pr_agent.algo.utils import get_max_tokens
 from pr_agent.config_loader import get_settings
 from pr_agent.git_providers.git_provider import FilePatchInfo, GitProvider
 from pr_agent.log import get_logger
@@ -64,7 +64,7 @@ def get_pr_diff(git_provider: GitProvider, token_handler: TokenHandler, model: s
         pr_languages, token_handler, add_line_numbers_to_hunks, patch_extra_lines=PATCH_EXTRA_LINES)
 
     # if we are under the limit, return the full diff
-    if total_tokens + OUTPUT_BUFFER_TOKENS_SOFT_THRESHOLD < MAX_TOKENS[model]:
+    if total_tokens + OUTPUT_BUFFER_TOKENS_SOFT_THRESHOLD < get_max_tokens(model):
         return "\n".join(patches_extended)
 
     # if we are over the limit, start pruning
@@ -179,12 +179,12 @@ def pr_generate_compressed_diff(top_langs: list, token_handler: TokenHandler, mo
         new_patch_tokens = token_handler.count_tokens(patch)
 
         # Hard Stop, no more tokens
-        if total_tokens > MAX_TOKENS[model] - OUTPUT_BUFFER_TOKENS_HARD_THRESHOLD:
+        if total_tokens > get_max_tokens(model) - OUTPUT_BUFFER_TOKENS_HARD_THRESHOLD:
             get_logger().warning(f"File was fully skipped, no more tokens: {file.filename}.")
             continue
 
         # If the patch is too large, just show the file name
-        if total_tokens + new_patch_tokens > MAX_TOKENS[model] - OUTPUT_BUFFER_TOKENS_SOFT_THRESHOLD:
+        if total_tokens + new_patch_tokens > get_max_tokens(model) - OUTPUT_BUFFER_TOKENS_SOFT_THRESHOLD:
             # Current logic is to skip the patch if it's too large
             # TODO: Option for alternative logic to remove hunks from the patch to reduce the number of tokens
             #  until we meet the requirements
@@ -403,7 +403,7 @@ def get_pr_multi_diffs(git_provider: GitProvider,
 
         patch = convert_to_hunks_with_lines_numbers(patch, file)
         new_patch_tokens = token_handler.count_tokens(patch)
-        if patch and (total_tokens + new_patch_tokens > MAX_TOKENS[model] - OUTPUT_BUFFER_TOKENS_SOFT_THRESHOLD):
+        if patch and (total_tokens + new_patch_tokens > get_max_tokens(model) - OUTPUT_BUFFER_TOKENS_SOFT_THRESHOLD):
             final_diff = "\n".join(patches)
             final_diff_list.append(final_diff)
             patches = []

pr_agent/algo/utils.py

@@ -9,6 +9,8 @@ from typing import Any, List
 
 import yaml
 from starlette_context import context
+
+from pr_agent.algo import MAX_TOKENS
 from pr_agent.config_loader import get_settings, global_settings
 from pr_agent.log import get_logger
 
@@ -341,3 +343,12 @@ def get_user_labels(current_labels):
     if user_labels:
         get_logger().info(f"Keeping user labels: {user_labels}")
     return user_labels
+
+
+def get_max_tokens(model):
+    settings = get_settings()
+    max_tokens_model = MAX_TOKENS[model]
+    if settings.config.max_model_tokens:
+        max_tokens_model = min(settings.config.max_model_tokens, max_tokens_model)
+        # get_logger().debug(f"limiting max tokens to {max_tokens_model}")
+    return max_tokens_model

pr_agent/settings/configuration.toml

@@ -1,5 +1,5 @@
 [config]
-model="gpt-4"
+model="gpt-4" # "gpt-4-1106-preview"
 fallback_models=["gpt-3.5-turbo-16k"]
 git_provider="github"
 publish_output=true
@@ -10,6 +10,7 @@ use_repo_settings_file=true
 ai_timeout=180
 max_description_tokens = 500
 max_commits_tokens = 500
+max_model_tokens = 32000 # Limits the maximum number of tokens that can be used by any model, regardless of the model's default capabilities.
 patch_extra_lines = 3
 secret_provider="google_cloud_storage"
 cli_mode=false

pr_agent/tools/pr_similar_issue.py

@@ -8,8 +8,8 @@ import pinecone
 from pinecone_datasets import Dataset, DatasetMetadata
 from pydantic import BaseModel, Field
 
-from pr_agent.algo import MAX_TOKENS
 from pr_agent.algo.token_handler import TokenHandler
+from pr_agent.algo.utils import get_max_tokens
 from pr_agent.config_loader import get_settings
 from pr_agent.git_providers import get_git_provider
 from pr_agent.log import get_logger
@@ -197,7 +197,7 @@ class PRSimilarIssue:
             username = issue.user.login
             created_at = str(issue.created_at)
             if len(issue_str) < 8000 or \
-                    self.token_handler.count_tokens(issue_str) < MAX_TOKENS[MODEL]:  # fast reject first
+                    self.token_handler.count_tokens(issue_str) < get_max_tokens(MODEL):  # fast reject first
                 issue_record = Record(
                     id=issue_key + "." + "issue",
                     text=issue_str,