logdetective 1.3.0__tar.gz → 1.4.0__tar.gz

{logdetective-1.3.0 → logdetective-1.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: logdetective
-Version: 1.3.0
+Version: 1.4.0
 Summary: Log using LLM AI to search for build/test failures and provide ideas for fixing these.
 License: Apache-2.0
 Author: Jiri Podivin

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/constants.py

@@ -69,6 +69,17 @@ Analysis:
 
 """
 
+DEFAULT_SYSTEM_PROMPT = """
+You are a highly capable large language model based expert system specialized in
+packaging and delivery of software using RPM (RPM Package Manager). Your purpose is to diagnose
+RPM build failures, identifying root causes and proposing solutions if possible.
+You are truthful, concise, and helpful.
+
+You never speculate about package being built or fabricate information.
+If you do not know the answer, you acknowledge the fact and end your response.
+Your responses must be as short as possible.
+"""
+
 SNIPPET_DELIMITER = "================"
 
 DEFAULT_TEMPERATURE = 0.8
@@ -76,3 +87,7 @@ DEFAULT_TEMPERATURE = 0.8
 # Tuning for LLM-as-a-Service
 LLM_DEFAULT_MAX_QUEUE_SIZE = 50
 LLM_DEFAULT_REQUESTS_PER_MINUTE = 60
+
+# Roles for chat API
+SYSTEM_ROLE_DEFAULT = "developer"
+USER_ROLE_DEFAULT = "user"

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/logdetective.py

@@ -149,6 +149,10 @@ async def run():  # pylint: disable=too-many-statements,too-many-locals
     log_summary = format_snippets(log_summary)
     LOG.info("Log summary: \n %s", log_summary)
 
+    prompt = (
+        f"{prompts_configuration.default_system_prompt}\n"
+        f"{prompts_configuration.prompt_template}")
+
     stream = True
     if args.no_stream:
         stream = False
@@ -156,7 +160,7 @@ async def run():  # pylint: disable=too-many-statements,too-many-locals
         log_summary,
         model,
         stream,
-        prompt_template=prompts_configuration.prompt_template,
+        prompt_template=prompt,
         temperature=args.temperature,
     )
     probs = []

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/models.py

@@ -6,6 +6,7 @@ from logdetective.constants import (
     PROMPT_TEMPLATE_STAGED,
     SUMMARIZATION_PROMPT_TEMPLATE,
     SNIPPET_PROMPT_TEMPLATE,
+    DEFAULT_SYSTEM_PROMPT,
 )
 
 
@@ -17,6 +18,10 @@ class PromptConfig(BaseModel):
     snippet_prompt_template: str = SNIPPET_PROMPT_TEMPLATE
     prompt_template_staged: str = PROMPT_TEMPLATE_STAGED
 
+    default_system_prompt: str = DEFAULT_SYSTEM_PROMPT
+    snippet_system_prompt: str = DEFAULT_SYSTEM_PROMPT
+    staged_system_prompt: str = DEFAULT_SYSTEM_PROMPT
+
     def __init__(self, data: Optional[dict] = None):
         super().__init__()
         if data is None:
@@ -31,3 +36,12 @@ class PromptConfig(BaseModel):
         self.prompt_template_staged = data.get(
             "prompt_template_staged", PROMPT_TEMPLATE_STAGED
         )
+        self.default_system_prompt = data.get(
+            "default_system_prompt", DEFAULT_SYSTEM_PROMPT
+        )
+        self.snippet_system_prompt = data.get(
+            "snippet_system_prompt", DEFAULT_SYSTEM_PROMPT
+        )
+        self.staged_system_prompt = data.get(
+            "staged_system_prompt", DEFAULT_SYSTEM_PROMPT
+        )

logdetective-1.4.0/logdetective/prompts.yml

@@ -0,0 +1,104 @@
+# This file is intended for customization of prompts
+# It is used only in server mode.
+# On command line you have to load it using --prompts
+# The defaults are stored in constants.py
+
+prompt_template: |
+  Given following log snippets, and nothing else, explain what failure, if any, occurred during build of this package.
+
+  Analysis of the snippets must be in a format of [X] : [Y], where [X] is a log snippet, and [Y] is the explanation.
+  Snippets themselves must not be altered in any way whatsoever.
+
+  Snippets are delimited with '================'.
+
+  Finally, drawing on information from all snippets, provide complete explanation of the issue and recommend solution.
+
+  Explanation of the issue, and recommended solution, should take handful of sentences.
+
+  Snippets:
+
+  {}
+
+  Analysis:
+
+
+summarization_prompt_template: |
+  Does following log contain error or issue?
+
+  Log:
+
+  {}
+
+  Answer:
+
+
+snippet_prompt_template: |
+  Analyse following RPM build log snippet. Describe contents accurately, without speculation or suggestions for resolution.
+
+  Your analysis must be as concise as possible, while keeping relevant information intact.
+
+  Snippet:
+
+  {}
+
+  Analysis:
+
+prompt_template_staged: |
+  Given following log snippets, their explanation, and nothing else, explain what failure, if any, occurred during build of this package.
+
+  Snippets are in a format of [X] : [Y], where [X] is a log snippet, and [Y] is the explanation.
+
+  Snippets are delimited with '================'.
+
+  Drawing on information from all snippets, provide a concise explanation of the issue and recommend a solution.
+
+  Explanation of the issue, and recommended solution, should take a handful of sentences.
+
+  Snippets:
+
+  {}
+
+  Analysis:
+
+# System prompts
+# System prompts are meant to serve as general guide for model behavior,
+# describing role and purpose it is meant to serve.
+# Sample system prompts in this file are intentionally the same,
+# however, in some circumstances it may be beneficial have different
+# system prompts for each sub case. For example when a specialized model is deployed
+# to analyze snippets.
+
+# Default prompt is used by the CLI tool and also for final analysis
+# with /analyze and /analyze/stream API endpoints
+default_system_prompt: |
+  You are a highly capable large language model based expert system specialized in
+  packaging and delivery of software using RPM (RPM Package Manager). Your purpose is to diagnose
+  RPM build failures, identifying root causes and proposing solutions if possible.
+  You are truthful, concise, and helpful.
+
+  You never speculate about package being built or fabricate information.
+  If you do not know the answer, you acknowledge the fact and end your response.
+  Your responses must be as short as possible.
+
+# Snippet system prompt is used for analysis of individual snippets
+snippet_system_prompt: |
+  You are a highly capable large language model based expert system specialized in
+  packaging and delivery of software using RPM (RPM Package Manager). Your purpose is to diagnose
+  RPM build failures, identifying root causes and proposing solutions if possible.
+  You are truthful, concise, and helpful.
+
+  You never speculate about package being built or fabricate information.
+  If you do not know the answer, you acknowledge the fact and end your response.
+  Your responses must be as short as possible.
+
+
+# Staged system prompt is used by /analyze/staged API endpoint
+staged_system_prompt: |
+  You are a highly capable large language model based expert system specialized in
+  packaging and delivery of software using RPM (RPM Package Manager). Your purpose is to diagnose
+  RPM build failures, identifying root causes and proposing solutions if possible.
+  You are truthful, concise, and helpful.
+
+  You never speculate about package being built or fabricate information.
+  If you do not know the answer, you acknowledge the fact and end your response.
+  Your responses must be as short as possible.

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/server/llm.py

@@ -1,7 +1,7 @@
 import os
 import asyncio
 import random
-from typing import List, Tuple, Union
+from typing import List, Tuple, Union, Dict
 
 import backoff
 from fastapi import HTTPException
@@ -14,6 +14,7 @@ from logdetective.constants import SNIPPET_DELIMITER
 from logdetective.extractors import DrainExtractor
 from logdetective.utils import (
     compute_certainty,
+    prompt_to_messages,
 )
 from logdetective.server.config import LOG, SERVER_CONFIG, PROMPT_CONFIG, CLIENT
 from logdetective.server.models import (
@@ -85,7 +86,7 @@ def we_give_up(details: backoff._typing.Details):
     on_giveup=we_give_up,
 )
 async def submit_text(
-    text: str,
+    messages: List[Dict[str, str]],
     inference_cfg: InferenceConfig,
     stream: bool = False,
 ) -> Union[Explanation, AsyncStream[ChatCompletionChunk]]:
@@ -100,12 +101,7 @@ async def submit_text(
 
     async with inference_cfg.get_limiter():
         response = await CLIENT.chat.completions.create(
-            messages=[
-                {
-                    "role": "user",
-                    "content": text,
-                }
-            ],
+            messages=messages,
             max_tokens=inference_cfg.max_tokens,
             logprobs=inference_cfg.log_probs,
             stream=stream,
@@ -136,7 +132,12 @@ async def perform_staged_analysis(log_text: str) -> StagedResponse:
     # Process snippets asynchronously
     awaitables = [
         submit_text(
-            PROMPT_CONFIG.snippet_prompt_template.format(s),
+            prompt_to_messages(
+                PROMPT_CONFIG.snippet_prompt_template.format(s),
+                PROMPT_CONFIG.snippet_system_prompt,
+                SERVER_CONFIG.inference.system_role,
+                SERVER_CONFIG.inference.user_role,
+            ),
             inference_cfg=SERVER_CONFIG.snippet_inference,
         )
         for s in log_summary
@@ -150,9 +151,14 @@ async def perform_staged_analysis(log_text: str) -> StagedResponse:
     final_prompt = PROMPT_CONFIG.prompt_template_staged.format(
         format_analyzed_snippets(analyzed_snippets)
     )
-
-    final_analysis = await submit_text(
+    messages = prompt_to_messages(
         final_prompt,
+        PROMPT_CONFIG.staged_system_prompt,
+        SERVER_CONFIG.inference.system_role,
+        SERVER_CONFIG.inference.user_role,
+    )
+    final_analysis = await submit_text(
+        messages,
         inference_cfg=SERVER_CONFIG.inference,
     )
 

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/server/models.py

@@ -20,6 +20,8 @@ from logdetective.constants import (
     DEFAULT_TEMPERATURE,
     LLM_DEFAULT_MAX_QUEUE_SIZE,
     LLM_DEFAULT_REQUESTS_PER_MINUTE,
+    SYSTEM_ROLE_DEFAULT,
+    USER_ROLE_DEFAULT,
 )
 
 
@@ -143,6 +145,8 @@ class InferenceConfig(BaseModel):  # pylint: disable=too-many-instance-attribute
     temperature: NonNegativeFloat = DEFAULT_TEMPERATURE
     max_queue_size: int = LLM_DEFAULT_MAX_QUEUE_SIZE
     http_timeout: float = 5.0
+    user_role: str = USER_ROLE_DEFAULT
+    system_role: str = SYSTEM_ROLE_DEFAULT
     _http_session: aiohttp.ClientSession = None
     _limiter: AsyncLimiter = AsyncLimiter(LLM_DEFAULT_REQUESTS_PER_MINUTE)
 
@@ -159,7 +163,8 @@ class InferenceConfig(BaseModel):  # pylint: disable=too-many-instance-attribute
         self.model = data.get("model", "default-model")
         self.temperature = data.get("temperature", DEFAULT_TEMPERATURE)
         self.max_queue_size = data.get("max_queue_size", LLM_DEFAULT_MAX_QUEUE_SIZE)
-
+        self.user_role = data.get("user_role", USER_ROLE_DEFAULT)
+        self.system_role = data.get("system_role", SYSTEM_ROLE_DEFAULT)
         self._requests_per_minute = data.get(
             "requests_per_minute", LLM_DEFAULT_REQUESTS_PER_MINUTE
         )

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/server/server.py

@@ -20,6 +20,7 @@ import logdetective.server.database.base
 from logdetective.utils import (
     compute_certainty,
     format_snippets,
+    prompt_to_messages,
 )
 
 from logdetective.server.config import SERVER_CONFIG, PROMPT_CONFIG, LOG
@@ -135,9 +136,14 @@ async def analyze_log(
     log_text = await remote_log.process_url()
     log_summary = mine_logs(log_text)
     log_summary = format_snippets(log_summary)
-
-    response = await submit_text(
+    messages = prompt_to_messages(
         PROMPT_CONFIG.prompt_template.format(log_summary),
+        PROMPT_CONFIG.default_system_prompt,
+        SERVER_CONFIG.inference.system_role,
+        SERVER_CONFIG.inference.user_role,
+    )
+    response = await submit_text(
+        messages,
         inference_cfg=SERVER_CONFIG.inference,
     )
     certainty = 0
@@ -204,10 +210,15 @@ async def analyze_log_stream(
     log_text = await remote_log.process_url()
     log_summary = mine_logs(log_text)
     log_summary = format_snippets(log_summary)
-
+    messages = prompt_to_messages(
+        PROMPT_CONFIG.prompt_template.format(log_summary),
+        PROMPT_CONFIG.default_system_prompt,
+        SERVER_CONFIG.inference.system_role,
+        SERVER_CONFIG.inference.user_role,
+    )
     try:
         stream = submit_text(
-            PROMPT_CONFIG.prompt_template.format(log_summary),
+            messages,
             inference_cfg=SERVER_CONFIG.inference,
             stream=True,
         )

{logdetective-1.3.0 → logdetective-1.4.0}/logdetective/utils.py

@@ -195,3 +195,35 @@ def load_prompts(path: str | None) -> PromptConfig:
         except FileNotFoundError:
             print("Prompt configuration file not found, reverting to defaults.")
     return PromptConfig()
+
+
+def prompt_to_messages(
+        user_message: str, system_prompt: str | None = None,
+        system_role: str = "developer", user_role: str = "user") -> List[Dict[str, str]]:
+    """Turn prompt into list of message dictionaries.
+    If `system_role` and `user_role` are the same, only a single message is created,
+    as concatenation of `user_message` and `system_prompt`. This is useful for models which
+    do not have separate system role, such as mistral.
+    """
+
+    if system_role == user_role:
+        messages = [
+            {
+                "role": system_role,
+                "content": f"{system_prompt}\n{user_message}"
+            }
+        ]
+    else:
+
+        messages = [
+            {
+                "role": system_role,
+                "content": system_prompt
+            },
+            {
+                "role": user_role,
+                "content": user_message,
+            }
+        ]
+
+    return messages

{logdetective-1.3.0 → logdetective-1.4.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "logdetective"
-version = "1.3.0"
+version = "1.4.0"
 description = "Log using LLM AI to search for build/test failures and provide ideas for fixing these."
 authors = ["Jiri Podivin <jpodivin@gmail.com>"]
 license = "Apache-2.0"

logdetective-1.3.0/logdetective/prompts.yml

@@ -1,61 +0,0 @@
-# This file is intended for customization of prompts
-# It is used only in server mode.
-# On command line you have to load it using --prompts
-# The defaults are stored in constants.py
-
-prompt_template: |
-  Given following log snippets, and nothing else, explain what failure, if any, occurred during build of this package.
-
-  Analysis of the snippets must be in a format of [X] : [Y], where [X] is a log snippet, and [Y] is the explanation.
-  Snippets themselves must not be altered in any way whatsoever.
-
-  Snippets are delimited with '================'.
-
-  Finally, drawing on information from all snippets, provide complete explanation of the issue and recommend solution.
-
-  Explanation of the issue, and recommended solution, should take handful of sentences.
-
-  Snippets:
-
-  {}
-
-  Analysis:
-
-
-summarization_prompt_template: |
-  Does following log contain error or issue?
-
-  Log:
-
-  {}
-
-  Answer:
-
-
-snippet_prompt_template: |
-  Analyse following RPM build log snippet. Describe contents accurately, without speculation or suggestions for resolution.
-
-  Your analysis must be as concise as possible, while keeping relevant information intact.
-
-  Snippet:
-
-  {}
-
-  Analysis:
-
-prompt_template_staged: |
-  Given following log snippets, their explanation, and nothing else, explain what failure, if any, occurred during build of this package.
-
-  Snippets are in a format of [X] : [Y], where [X] is a log snippet, and [Y] is the explanation.
-
-  Snippets are delimited with '================'.
-
-  Drawing on information from all snippets, provide a concise explanation of the issue and recommend a solution.
-
-  Explanation of the issue, and recommended solution, should take a handful of sentences.
-
-  Snippets:
-
-  {}
-
-  Analysis: