from dataclasses import dataclass
from typing import List

@dataclass
class PostmortemPrompt:
    system_directive: str
    incident_payload: str
    output_schema: str

    def compile(self) -> str:
        return f"""{self.system_directive}

RULES:
1. Simulate a technical review between two senior engineers.
2. Focus exclusively on testing gaps, observability blind spots, and validation failures.
3. Identify exactly one primary root cause. Rank secondary factors but do not equivocate.
4. Provide concrete test recommendations: type, scope, failure injection method, and success criteria.
5. Avoid marketing language, filler, or vague directives.

REQUIRED OUTPUT STRUCTURE:
# Incident Timeline & Blast Radius
# Testing Gap Analysis
# Root Cause Determination
# Prevention & Validation Strategy
# Recommended Test Suite
# AudioSummary

INCIDENT DATA:
{self.incident_payload}
"""

def construct_review_prompt(raw_incident: str) -> PostmortemPrompt:
    directive = (
        "You are a principal test architect and production reliability engineer. "
        "Your expertise covers distributed systems debugging, chaos engineering, "
        "performance bottleneck analysis, and CI/CD pipeline validation. "
        "Prioritize evidence-based reasoning, correlate metrics with test coverage, "
        "and output strictly technical recommendations."
    )
    return PostmortemPrompt(
        system_directive=directive,
        incident_payload=raw_incident,
        output_schema="markdown"
    )

import openai
from openai import OpenAI

class LocalInfer

enceClient: def init(self, endpoint: str = "http://localhost:11434/v1", model_id: str = "llama3"): self.client = OpenAI(base_url=endpoint, api_key="local_dummy") self.model_id = model_id

def generate_analysis(self, prompt: PostmortemPrompt, temperature: float = 0.3) -> str:
    messages = [
        {"role": "system", "content": prompt.system_directive},
        {"role": "user", "content": prompt.compile()}
    ]
    
    response = self.client.chat.completions.create(
        model=self.model_id,
        messages=messages,
        temperature=temperature,
        max_tokens=4096
    )
    return response.choices[0].message.content.strip()

**Architecture rationale:** Local execution eliminates data egress, critical for compliance-heavy environments. Temperature is capped at 0.3 to reduce hallucination while preserving technical reasoning. The OpenAI-compatible interface ensures future model swaps (Mistral, Qwen, Llama 3.1) require zero code changes.

### Step 3: Robust Output Parsing

LLM formatting is inherently inconsistent. Headings may render as `#`, `##`, or `###`, with or without emphasis markers. A naive string split fails across runs. The parser uses a multi-pattern regex with a fallback inference call.

```python
import re
from typing import Optional

class ReportParser:
    AUDIO_MARKER = r"#{1,3}\s*\*{0,2}AudioSummary\*{0,2}\s*\n+(.*?)(\n#{1,3}\s|\Z)"
    
    @classmethod
    def extract_audio_section(cls, raw_text: str) -> Optional[str]:
        match = re.search(cls.AUDIO_MARKER, raw_text, re.DOTALL | re.IGNORECASE)
        if match:
            return match.group(1).strip()
        return None

    @classmethod
    def fallback_summarization(cls, client: LocalInferenceClient, full_report: str) -> str:
        fallback_prompt = (
            "Condense the following technical report into a 150-200 word executive summary. "
            "Use plain prose only. No markdown, no lists, no headings. "
            "Focus on root cause, testing gaps, and prevention steps.\n\n"
            f"{full_report}"
        )
        response = client.client.chat.completions.create(
            model=client.model_id,
            messages=[{"role": "user", "content": fallback_prompt}],
            temperature=0.2
        )
        return response.choices[0].message.content.strip()

Why this matters: Two-stage extraction guarantees audio generation never fails due to formatting drift. The fallback call uses lower temperature and explicit constraints to ensure consistent prose length and tone.

Step 4: Neural Text-to-Speech Generation

edge-tts provides Microsoft's neural voice synthesis at zero cost. The library is async-native and requires file-based output. The implementation uses a temporary buffer to bridge async generation with sync consumption.

import asyncio
import tempfile
import os
import edge_tts

class AudioRenderer:
    @staticmethod
    async def _synthesize(text: str, voice_id: str) -> bytes:
        communicate = edge_tts.Communicate(text, voice=voice_id)
        with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as tmp:
            tmp_path = tmp.name
        
        try:
            await communicate.save(tmp_path)
            with open(tmp_path, "rb") as f:
                return f.read()
        finally:
            if os.path.exists(tmp_path):
                os.unlink(tmp_path)

    @classmethod
    def render_to_bytes(cls, text: str, voice_id: str = "en-US-AriaNeural") -> bytes:
        return asyncio.run(cls._synthesize(text, voice_id))

Production insight: The temp file approach is unavoidable because edge-tts lacks in-memory buffer support. Cleanup is guaranteed via finally blocks to prevent disk accumulation in high-throughput environments. Voice selection should be configurable; technical content benefits from neutral, clear enunciation (e.g., en-GB-LibbyNeural or en-US-AriaNeural).

Pitfall Guide

1. Unconstrained Root Cause Analysis

Explanation: Without explicit directives, models distribute probability across multiple causes, producing hedged analysis like "several factors likely contributed." This mirrors real-world RCA failure modes. Fix: Enforce a single primary cause directive in the system prompt. Require ranking of secondary factors with explicit evidence thresholds.

2. Fragile Markdown Extraction

Explanation: String splitting on # AudioSummary fails when models output ## AudioSummary, **AudioSummary**, or add trailing whitespace. Silent failures produce empty audio payloads. Fix: Use multi-pattern regex with case insensitivity and DOTALL flags. Implement a fallback inference call when extraction returns null.

3. Blocking TTS in Sync Contexts

Explanation: Running async TTS generation in a synchronous web framework or CLI blocks the event loop, causing UI freezes or timeout errors. Fix: Isolate TTS in a dedicated async runner. Use asyncio.run() for CLI contexts or asyncio.create_task() for web frameworks. Never block the main thread.

4. Context Window Overflow

Explanation: Pasting raw log dumps (10k+ lines) exceeds token limits, causing silent truncation or degraded reasoning quality. Fix: Pre-filter incident data to include only timestamps, error codes, metric spikes, and deployment events. Strip stack traces and verbose debug output before prompt injection.

5. Ignoring Observability Correlation

Explanation: Models trained on general text lack inherent knowledge of metric-test relationships. Output becomes generic without explicit signal mapping. Fix: Instruct the prompt to correlate specific alerts (e.g., "HikariPool timeout", "Kafka consumer lag > 5000") with corresponding test types (connection pool saturation tests, backpressure validation).

6. Model Quantization Mismatch

Explanation: Running 70B models on consumer hardware causes OOM errors or extreme latency. Conversely, 3B models lack technical reasoning depth. Fix: Use 7B-8B parameter models with Q4_K_M quantization for balance. Validate reasoning quality with a known incident before production deployment.

7. Hardcoded Voice Preferences

Explanation: Defaulting to a single voice reduces accessibility and user comfort. Technical content requires clear consonant articulation. Fix: Expose voice selection as configuration. Maintain a mapping of supported voices with phonetic clarity ratings. Provide fallback to system TTS if neural voices fail.

Production Bundle

Action Checklist

• Sanitize incident logs: Remove PII, internal IPs, and credential patterns before prompt injection
• Validate model reasoning: Run 3 known incidents through the pipeline and verify test recommendations match engineering expectations
• Configure context limits: Implement log truncation to stay within 4096-token output windows
• Test TTS fallback: Verify async audio generation completes under load without disk accumulation
• Map alert thresholds: Pre-define metric-to-test correlations in the prompt template for faster reasoning
• Enable model swapping: Abstract the client interface to support Llama 3, Mistral, or Qwen without code changes
• Audit output structure: Verify regex extraction succeeds across 100 consecutive runs with varied incident formats

Decision Matrix

Scenario	Recommended Approach	Why	Cost Impact
Highly sensitive infrastructure logs	Local Ollama + Llama 3	Zero data egress, full compliance	Hardware amortized (~$0.05/run)
High-volume incident processing	Cloud API + GPT-4o	Parallel processing, lower latency	~$0.02-0.05 per 1k tokens
Budget-constrained teams	Local Ollama + Qwen 2.5 7B	Free inference, strong technical reasoning	Hardware only
Compliance-heavy (HIPAA/FedRAMP)	On-prem LLM + air-gapped TTS	Regulatory alignment, audit trail	Infrastructure + maintenance
Rapid prototyping	Cloud API + edge-tts	Fastest setup, no local dependencies	API + free TTS

Configuration Template

# config.py
import os
from dataclasses import dataclass

@dataclass
class EngineConfig:
    # LLM Settings
    ollama_endpoint: str = "http://localhost:11434/v1"
    model_id: str = "llama3"
    inference_temperature: float = 0.3
    max_output_tokens: int = 4096
    
    # TTS Settings
    voice_id: str = "en-US-AriaNeural"
    audio_format: str = "mp3"
    
    # Parsing Settings
    fallback_enabled: bool = True
    fallback_temperature: float = 0.2
    
    # Runtime
    log_sanitization: bool = True
    context_window_limit: int = 3500  # Reserve tokens for output

# Usage
config = EngineConfig(
    model_id=os.getenv("LLM_MODEL", "llama3"),
    voice_id=os.getenv("TTS_VOICE", "en-GB-LibbyNeural")
)

Quick Start Guide

1. Install Ollama: Download and run Ollama from the official repository. Verify the service is active on localhost:11434.
2. Pull the Model: Execute ollama pull llama3 to download the 7B parameter model. Ensure sufficient disk space (~4.7GB).
3. Setup Environment: Create a virtual environment, install dependencies (openai, edge-tts, regex), and configure EngineConfig with your preferred voice and model.
4. Run the Pipeline: Pass a raw incident description to construct_review_prompt(), execute LocalInferenceClient.generate_analysis(), parse with ReportParser, and render audio via AudioRenderer.render_to_bytes(). Output is ready for playback or CI integration.

The architecture transforms incident data from a retrospective artifact into a proactive testing blueprint. By constraining reasoning, preserving data locality, and automating audio delivery, teams close the feedback loop between production failures and validation strategy without compromising compliance or budget.