logdetective 0.2.6__tar.gz → 0.2.8__tar.gz

{logdetective-0.2.6 → logdetective-0.2.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: logdetective
-Version: 0.2.6
+Version: 0.2.8
 Summary: Log using LLM AI to search for build/test failures and provide ideas for fixing these.
 License: Apache-2.0
 Author: Jiri Podivin
@@ -14,14 +14,16 @@ Classifier: Natural Language :: English
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
 Classifier: Topic :: Internet :: Log Analysis
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development :: Debuggers
 Provides-Extra: server
 Requires-Dist: drain3 (>=0.9.11,<0.10.0)
 Requires-Dist: huggingface-hub (>0.23.2)
-Requires-Dist: llama-cpp-python (>=0.2.56,<0.3.0,!=0.2.86)
-Requires-Dist: requests (>=2.31.0,<3.0.0)
+Requires-Dist: llama-cpp-python (>0.2.56,!=0.2.86)
+Requires-Dist: numpy (>=1.26.0,<2.0.0)
+Requires-Dist: requests (>0.2.31)
 Project-URL: homepage, https://github.com/fedora-copr/logdetective
 Project-URL: issues, https://github.com/fedora-copr/logdetective/issues
 Description-Content-Type: text/markdown
@@ -38,7 +40,7 @@ A Python tool to analyze logs using a Language Model (LLM) and Drain template mi
 Installation
 ------------
 
-** Fedora 40+ **
+**Fedora 40+**
 
     dnf install logdetective
 
@@ -193,6 +195,11 @@ Requests can then be made with post requests, for example:
 
     curl --header "Content-Type: application/json" --request POST --data '{"url":"<YOUR_URL_HERE>"}' http://localhost:8080/analyze
 
+For more accurate responses, you can use `/analyze/staged` endpoint. This will submit snippets to model for individual analysis first.
+Afterwards the model outputs are used to construct final prompt. This will take substantially longer, compared to plain `/analyze`
+
+    curl --header "Content-Type: application/json" --request POST --data '{"url":"<YOUR_URL_HERE>"}' http://localhost:8080/analyze/staged
+
 We also have a Containerfile and composefile to run the logdetective server and llama server in containers.
 
 Before doing `podman-compose up`, make sure to set `MODELS_PATH` environment variable and point to a directory with your local model files:
@@ -204,6 +211,11 @@ $ ll $MODELS_PATH
 
 If the variable is not set, `./models` is mounted inside by default.
 
+Model can be downloaded from [our Hugging Space](https://huggingface.co/fedora-copr) by:
+```
+$ curl -L -o models/mistral-7b-instruct-v0.2.Q4_K_S.gguf https://huggingface.co/fedora-copr/Mistral-7B-Instruct-v0.2-GGUF/resolve/main/ggml-model-Q4_K_S.gguf
+```
+
 
 License
 -------

{logdetective-0.2.6 → logdetective-0.2.8}/README.md

@@ -10,7 +10,7 @@ A Python tool to analyze logs using a Language Model (LLM) and Drain template mi
 Installation
 ------------
 
-** Fedora 40+ **
+**Fedora 40+**
 
     dnf install logdetective
 
@@ -165,6 +165,11 @@ Requests can then be made with post requests, for example:
 
     curl --header "Content-Type: application/json" --request POST --data '{"url":"<YOUR_URL_HERE>"}' http://localhost:8080/analyze
 
+For more accurate responses, you can use `/analyze/staged` endpoint. This will submit snippets to model for individual analysis first.
+Afterwards the model outputs are used to construct final prompt. This will take substantially longer, compared to plain `/analyze`
+
+    curl --header "Content-Type: application/json" --request POST --data '{"url":"<YOUR_URL_HERE>"}' http://localhost:8080/analyze/staged
+
 We also have a Containerfile and composefile to run the logdetective server and llama server in containers.
 
 Before doing `podman-compose up`, make sure to set `MODELS_PATH` environment variable and point to a directory with your local model files:
@@ -176,6 +181,11 @@ $ ll $MODELS_PATH
 
 If the variable is not set, `./models` is mounted inside by default.
 
+Model can be downloaded from [our Hugging Space](https://huggingface.co/fedora-copr) by:
+```
+$ curl -L -o models/mistral-7b-instruct-v0.2.Q4_K_S.gguf https://huggingface.co/fedora-copr/Mistral-7B-Instruct-v0.2-GGUF/resolve/main/ggml-model-Q4_K_S.gguf
+```
+
 
 License
 -------

{logdetective-0.2.6 → logdetective-0.2.8}/logdetective/constants.py

@@ -30,3 +30,16 @@ Log:
 Answer:
 
 """
+
+SNIPPET_PROMPT_TEMPLATE = """
+Analyse following RPM build log snippet.
+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.
+
+Snippet:
+
+{}
+
+Analysis:
+
+"""

{logdetective-0.2.6 → logdetective-0.2.8}/logdetective/logdetective.py

@@ -3,7 +3,8 @@ import logging
 import sys
 
 from logdetective.constants import DEFAULT_ADVISOR
-from logdetective.utils import process_log, initialize_model, retrieve_log_content, format_snippets
+from logdetective.utils import (
+    process_log, initialize_model, retrieve_log_content, format_snippets, compute_certainty)
 from logdetective.extractors import LLMExtractor, DrainExtractor
 
 LOG = logging.getLogger("logdetective")
@@ -92,15 +93,21 @@ def main():
     if args.no_stream:
         stream = False
     response = process_log(log_summary, model, stream)
+    probs = []
     print("Explanation:")
     if args.no_stream:
         print(response["choices"][0]["text"])
+        probs = response["choices"][0]["logprobs"]["top_logprobs"]
     else:
         # Stream the output
         for chunk in response:
+            if isinstance(chunk["choices"][0]["logprobs"], dict):
+                probs.extend(chunk["choices"][0]["logprobs"]["top_logprobs"])
             delta = chunk['choices'][0]['text']
             print(delta, end='', flush=True)
-        print()
+    certainty = compute_certainty(probs)
+
+    print(f"\nResponse certainty: {certainty:.2f}%\n")
 
 
 if __name__ == "__main__":

logdetective-0.2.8/logdetective/server.py

@@ -0,0 +1,242 @@
+import asyncio
+import json
+import logging
+import os
+from typing import List, Annotated
+
+from llama_cpp import CreateCompletionResponse
+from fastapi import FastAPI, HTTPException, Depends, Header
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+import requests
+
+from logdetective.constants import PROMPT_TEMPLATE, SNIPPET_PROMPT_TEMPLATE
+from logdetective.extractors import DrainExtractor
+from logdetective.utils import validate_url, compute_certainty
+
+class BuildLog(BaseModel):
+    """Model of data submitted to API.
+    """
+    url: str
+
+
+class Response(BaseModel):
+    """Model of data returned by Log Detective API
+
+    explanation: CreateCompletionResponse
+        https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.llama_types.CreateCompletionResponse
+    response_certainty: float
+    """
+    explanation: CreateCompletionResponse
+    response_certainty: float
+
+
+class StagedResponse(Response):
+    """Model of data returned by Log Detective API when called when staged response
+    is requested. Contains list of reponses to prompts for individual snippets.
+
+    explanation: CreateCompletionResponse
+        https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.llama_types.CreateCompletionResponse
+    response_certainty: float
+    snippets: list of CreateCompletionResponse
+    """
+    snippets: List[CreateCompletionResponse]
+
+
+LOG = logging.getLogger("logdetective")
+
+
+LLM_CPP_HOST = os.environ.get("LLAMA_CPP_HOST", "localhost")
+LLM_CPP_SERVER_ADDRESS = f"http://{LLM_CPP_HOST}"
+LLM_CPP_SERVER_PORT = os.environ.get("LLAMA_CPP_SERVER_PORT", 8000)
+LLM_CPP_SERVER_TIMEOUT = os.environ.get("LLAMA_CPP_SERVER_TIMEOUT", 600)
+LOG_SOURCE_REQUEST_TIMEOUT = os.environ.get("LOG_SOURCE_REQUEST_TIMEOUT", 60)
+API_TOKEN = os.environ.get("LOGDETECTIVE_TOKEN", None)
+
+def requires_token_when_set(authentication: Annotated[str | None, Header()] = None):
+    """
+    FastAPI Depend function that expects a header named Authentication
+
+    If LOGDETECTIVE_TOKEN env var is set, validate the client-supplied token
+    otherwise ignore it
+    """
+    if not API_TOKEN:
+        LOG.info("LOGDETECTIVE_TOKEN env var not set, authentication disabled")
+        # no token required, means local dev environment
+        return
+    token = None
+    if authentication:
+        try:
+            token = authentication.split(" ", 1)[1]
+        except (ValueError, IndexError):
+            LOG.warning(
+                "Authentication header has invalid structure (%s), it should be 'Bearer TOKEN'",
+                authentication)
+            # eat the exception and raise 401 below
+            token = None
+        if token == API_TOKEN:
+            return
+    LOG.info("LOGDETECTIVE_TOKEN env var is set (%s), clien token = %s",
+             API_TOKEN, token)
+    raise HTTPException(status_code=401, detail=f"Token {token} not valid.")
+
+app = FastAPI(dependencies=[Depends(requires_token_when_set)])
+
+
+def process_url(url: str) -> str:
+    """Validate log URL and return log text.
+    """
+    if validate_url(url=url):
+        try:
+            log_request = requests.get(url, timeout=int(LOG_SOURCE_REQUEST_TIMEOUT))
+        except requests.RequestException as ex:
+            raise HTTPException(
+                status_code=400,
+                detail=f"We couldn't obtain the logs: {ex}") from ex
+
+        if not log_request.ok:
+            raise HTTPException(status_code=400,
+                                detail="Something went wrong while getting the logs: "
+                                    f"[{log_request.status_code}] {log_request.text}")
+    else:
+        LOG.error("Invalid URL received ")
+        raise HTTPException(status_code=400,
+                            detail=f"Invalid log URL: {url}")
+
+    return log_request.text
+
+
+def mine_logs(log: str) -> List[str]:
+    """Extract snippets from log text
+    """
+    extractor = DrainExtractor(verbose=True, context=True, max_clusters=8)
+
+    LOG.info("Getting summary")
+    log_summary = extractor(log)
+
+    ratio = len(log_summary) / len(log.split('\n'))
+    LOG.debug("Log summary: \n %s", log_summary)
+    LOG.info("Compression ratio: %s", ratio)
+
+
+    return log_summary
+
+async def submit_text(text: str, max_tokens: int = 0, log_probs: int = 1, stream: bool = False,
+                      model: str = "default-model"):
+    """Submit prompt to LLM.
+    max_tokens: number of tokens to be produces, 0 indicates run until encountering EOS
+    log_probs: number of token choices to produce log probs for
+    """
+    LOG.info("Analyzing the text")
+    data = {
+            "prompt": text,
+            "max_tokens": str(max_tokens),
+            "logprobs": str(log_probs),
+            "stream": stream,
+            "model": model}
+
+    try:
+        # Expects llama-cpp server to run on LLM_CPP_SERVER_ADDRESS:LLM_CPP_SERVER_PORT
+        response = requests.post(
+            f"{LLM_CPP_SERVER_ADDRESS}:{LLM_CPP_SERVER_PORT}/v1/completions",
+            headers={"Content-Type":"application/json"},
+            data=json.dumps(data),
+            timeout=int(LLM_CPP_SERVER_TIMEOUT),
+            stream=stream)
+    except requests.RequestException as ex:
+        raise HTTPException(
+            status_code=400,
+            detail=f"Llama-cpp query failed: {ex}") from ex
+    if not stream:
+        if not response.ok:
+            raise HTTPException(
+                status_code=400,
+                detail="Something went wrong while getting a response from the llama server: "
+                    f"[{response.status_code}] {response.text}")
+        try:
+            response = json.loads(response.text)
+        except UnicodeDecodeError as ex:
+            LOG.error("Error encountered while parsing llama server response: %s", ex)
+            raise HTTPException(
+                status_code=400,
+                detail=f"Couldn't parse the response.\nError: {ex}\nData: {response.text}") from ex
+    else:
+        return response
+
+    return CreateCompletionResponse(response)
+
+
+@app.post("/analyze", response_model=Response)
+async def analyze_log(build_log: BuildLog):
+    """Provide endpoint for log file submission and analysis.
+    Request must be in form {"url":"<YOUR_URL_HERE>"}.
+    URL must be valid for the request to be passed to the LLM server.
+    Meaning that it must contain appropriate scheme, path and netloc,
+    while lacking  result, params or query fields.
+    """
+    log_text = process_url(build_log.url)
+    log_summary = mine_logs(log_text)
+    response = await submit_text(PROMPT_TEMPLATE.format(log_summary))
+    certainty = 0
+
+    if "logprobs" in response["choices"][0]:
+        try:
+            certainty = compute_certainty(
+                response["choices"][0]["logprobs"]["top_logprobs"])
+        except ValueError as ex:
+            LOG.error("Error encountered while computing certainty: %s", ex)
+            raise HTTPException(
+                status_code=400,
+                detail=f"Couldn't compute certainty with data:\n"
+                f"{response["choices"][0]["logprobs"]["top_logprobs"]}") from ex
+
+    return Response(explanation=response, response_certainty=certainty)
+
+
+@app.post("/analyze/staged", response_model=StagedResponse)
+async def analyze_log_staged(build_log: BuildLog):
+    """Provide endpoint for log file submission and analysis.
+    Request must be in form {"url":"<YOUR_URL_HERE>"}.
+    URL must be valid for the request to be passed to the LLM server.
+    Meaning that it must contain appropriate scheme, path and netloc,
+    while lacking  result, params or query fields.
+    """
+    log_text = process_url(build_log.url)
+    log_summary = mine_logs(log_text)
+
+    # Process snippets asynchronously
+    analyzed_snippets = await asyncio.gather(
+        *[submit_text(SNIPPET_PROMPT_TEMPLATE.format(s)) for s in log_summary])
+
+    final_analysis = await submit_text(
+        PROMPT_TEMPLATE.format([e["choices"][0]["text"] for e in analyzed_snippets]))
+
+    certainty = 0
+    if "logprobs" in final_analysis["choices"][0]:
+        try:
+            certainty = compute_certainty(
+                final_analysis["choices"][0]["logprobs"]["top_logprobs"])
+        except ValueError as ex:
+            LOG.error("Error encountered while computing certainty: %s", ex)
+            raise HTTPException(
+                status_code=400,
+                detail=f"Couldn't compute certainty with data:\n"
+                f"{final_analysis["choices"][0]["logprobs"]["top_logprobs"]}") from ex
+
+    return StagedResponse(
+        explanation=final_analysis, snippets=analyzed_snippets, response_certainty=certainty)
+
+
+@app.post("/analyze/stream", response_class=StreamingResponse)
+async def analyze_log_stream(build_log: BuildLog):
+    """Stream response endpoint for Logdetective.
+    Request must be in form {"url":"<YOUR_URL_HERE>"}.
+    URL must be valid for the request to be passed to the LLM server.
+    Meaning that it must contain appropriate scheme, path and netloc,
+    while lacking  result, params or query fields.
+    """
+    log_text = process_url(build_log.url)
+    log_summary = mine_logs(log_text)
+    stream = await submit_text(PROMPT_TEMPLATE.format(log_summary), stream=True)
+
+    return StreamingResponse(stream)

{logdetective-0.2.6 → logdetective-0.2.8}/logdetective/utils.py

@@ -1,10 +1,11 @@
 import logging
 import os
+from typing import Iterator, List, Dict
 from urllib.parse import urlparse
-
+import numpy as np
 import requests
 
-from llama_cpp import Llama
+from llama_cpp import Llama, CreateCompletionResponse, CreateCompletionStreamResponse
 from logdetective.constants import PROMPT_TEMPLATE
 
 
@@ -60,20 +61,38 @@ def initialize_model(model_pth: str, filename_suffix: str = ".gguf", verbose: bo
         model = Llama(
             model_path=model_pth,
             n_ctx=0,  # Maximum context for the model
-            verbose=verbose)
+            verbose=verbose,
+            logits_all=True)
     else:
         model = Llama.from_pretrained(
             model_pth,
             f"*{filename_suffix}",
             n_ctx=0,  # Maximum context for the model
-            verbose=verbose)
+            verbose=verbose,
+            logits_all=True)
 
     return model
 
 
-def process_log(log: str, model: Llama, stream: bool) -> str:
+def compute_certainty(probs: List[Dict[str, float] | None]) -> float:
+    """Compute certainty of repsponse based on average logit probability.
+    Log probability is log(p), isn't really readable for most people, especially in compound.
+    In this case it's just a matter of applying inverse operation exp.
+    Of course that leaves you with a value in range <0, 1> so it needs to be multiplied by 100.
+    Simply put, this is the most straightforward way to get the numbers out.
     """
-    Processes a given log using the provided language model and returns its summary.
+
+    top_logprobs = [
+        np.exp(x) * 100 for e in probs if isinstance(e, dict) for x in e.values()]
+    certainty = np.median(top_logprobs, axis=0)
+    if np.isnan(certainty):
+        raise ValueError("NaN certainty of answer")
+    return certainty
+
+
+def process_log(log: str, model: Llama, stream: bool) -> (
+        CreateCompletionResponse | Iterator[CreateCompletionStreamResponse]):
+    """Processes a given log using the provided language model and returns its summary.
 
     Args:
         log (str): The input log to be processed.
@@ -82,10 +101,13 @@ def process_log(log: str, model: Llama, stream: bool) -> str:
     Returns:
         str: The summary of the given log generated by the language model.
     """
-    return model(
+    response = model(
         prompt=PROMPT_TEMPLATE.format(log),
         stream=stream,
-        max_tokens=0)
+        max_tokens=0,
+        logprobs=1)
+
+    return response
 
 
 def retrieve_log_content(log_path: str) -> str:

{logdetective-0.2.6 → logdetective-0.2.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "logdetective"
-version = "0.2.6"
+version = "0.2.8"
 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"
@@ -27,10 +27,11 @@ issues = "https://github.com/fedora-copr/logdetective/issues"
 
 [tool.poetry.dependencies]
 python = "^3.11"
-requests = "^2.31.0"
-llama-cpp-python = "^0.2.56,!=0.2.86"
+requests = ">0.2.31"
+llama-cpp-python = ">0.2.56,!=0.2.86"
 drain3 = "^0.9.11"
 huggingface-hub = ">0.23.2"
+numpy = "^1.26.0"
 
 [build-system]
 requires = ["poetry-core"]

logdetective-0.2.6/logdetective/server.py

@@ -1,87 +0,0 @@
-import logging
-import os
-import json
-
-from fastapi import FastAPI, HTTPException
-from pydantic import BaseModel
-
-import requests
-
-from logdetective.constants import PROMPT_TEMPLATE
-from logdetective.extractors import DrainExtractor
-from logdetective.utils import validate_url
-
-class BuildLog(BaseModel):
-    """Model of data submitted to API.
-    """
-    url: str
-
-LOG = logging.getLogger("logdetective")
-
-app = FastAPI()
-
-LLM_CPP_HOST = os.environ.get("LLAMA_CPP_HOST", "localhost")
-LLM_CPP_SERVER_ADDRESS = f"http://{LLM_CPP_HOST}"
-LLM_CPP_SERVER_PORT = os.environ.get("LLAMA_CPP_SERVER_PORT", 8000)
-LLM_CPP_SERVER_TIMEOUT = os.environ.get("LLAMA_CPP_SERVER_TIMEOUT", 600)
-LOG_SOURCE_REQUEST_TIMEOUT = os.environ.get("LOG_SOURCE_REQUEST_TIMEOUT", 60)
-
-@app.post("/analyze", )
-async def analyze_log(build_log: BuildLog):
-    """Provide endpoint for log file submission and analysis.
-    Request must be in form {"url":"<YOUR_URL_HERE>"}.
-    URL must be valid for the request to be passed to the LLM server.
-    Meaning that it must contain appropriate scheme, path and netloc,
-    while lacking  result, params or query fields.
-    """
-    extractor = DrainExtractor(verbose=True, context=True, max_clusters=8)
-
-    LOG.info("Getting summary")
-    # Perform basic validation of the URL
-    if validate_url(url=build_log.url):
-        try:
-            log_request = requests.get(build_log.url, timeout=int(LOG_SOURCE_REQUEST_TIMEOUT))
-        except requests.RequestException as ex:
-            raise HTTPException(
-                status_code=400,
-                detail=f"We couldn't obtain the logs: {ex}") from ex
-
-        if not log_request.ok:
-            raise HTTPException(status_code=400,
-                                detail="Something went wrong while getting the logs: "
-                                    f"[{log_request.status_code}] {log_request.text}")
-    else:
-        LOG.error("Invalid URL received ")
-        raise HTTPException(status_code=400,
-                            detail=f"Invalid log URL: {build_log.url}")
-
-    log = log_request.text
-    log_summary = extractor(log)
-
-    ratio = len(log_summary) / len(log.split('\n'))
-    LOG.debug("Log summary: \n %s", log_summary)
-    LOG.info("Compression ratio: %s", ratio)
-
-    LOG.info("Analyzing the text")
-    data = {
-            "prompt": PROMPT_TEMPLATE.format(log_summary),
-            "max_tokens": "0"}
-
-    try:
-        # Expects llama-cpp server to run on LLM_CPP_SERVER_ADDRESS:LLM_CPP_SERVER_PORT
-        response = requests.post(
-            f"{LLM_CPP_SERVER_ADDRESS}:{LLM_CPP_SERVER_PORT}/v1/completions",
-            headers={"Content-Type":"application/json"},
-            data=json.dumps(data),
-            timeout=int(LLM_CPP_SERVER_TIMEOUT))
-    except requests.RequestException as ex:
-        raise HTTPException(
-            status_code=400,
-            detail=f"Llama-cpp query failed: {ex}") from ex
-
-    if not log_request.ok:
-        raise HTTPException(
-            status_code=400,
-            detail="Something went wrong while getting a response from the llama server: "
-                   f"[{log_request.status_code}] {log_request.text}")
-    return response.text