@agentix-e/log-parser-llm

Convenient OpenAI-compatible LLM provider for the log-parser framework -- implements ILLMProvider using the Vercel AI SDK with structured output.

Overview

@agentix-e/log-parser-llm provides the OpenAICompatibleProvider -- a single class that implements ILLMProvider for any OpenAI-compatible endpoint. It uses the Vercel AI SDK's generateObject() with a Zod schema to enforce structured NER-style template extraction. Every response is validated against the schema before being returned.

Supported backends:

| Provider | Default Endpoint | API Key Required | |----------|-----------------|:---:| | Ollama | http://localhost:11434/v1 | No | | OpenAI | https://api.openai.com/v1 | Yes | | DeepSeek | https://api.deepseek.com/v1 | Yes | | Anthropic | https://api.anthropic.com/v1 | Yes | | Custom | Any OpenAI-compatible URL | Depends |

No network calls occur at construction time -- the provider lazily creates the underlying AI SDK client. Network activity only occurs when extractTemplate() is called.

Installation

npm install @agentix-e/log-parser-llm @agentix-e/log-parser-core ai zod @ai-sdk/openai-compatible

Quick Start

Ollama (Local, Zero Cost)

# Start Ollama with a model
ollama pull qwen2.5:7b

import { OpenAICompatibleProvider } from '@agentix-e/log-parser-llm';
import { LogParserPipeline } from '@agentix-e/log-parser-core';

const llm = new OpenAICompatibleProvider({
  provider: 'ollama',
  model: 'qwen2.5:7b',
});

const pipeline = new LogParserPipeline({ llmProvider: llm });

const result = pipeline.parse('User alice uploaded file report_2025.pdf (2.4MB)');
console.log(result.template); // "User <*> uploaded file <*> (<*>MB)"
console.log(result.source);   // "llm"

OpenAI

import { OpenAICompatibleProvider } from '@agentix-e/log-parser-llm';

const llm = new OpenAICompatibleProvider({
  provider: 'openai',
  model: 'gpt-4o-mini',
  apiKey: process.env.OPENAI_API_KEY,
});

const pipeline = new LogParserPipeline({ llmProvider: llm });
// ~$0.0002 per call with gpt-4o-mini

DeepSeek

const llm = new OpenAICompatibleProvider({
  provider: 'deepseek',
  model: 'deepseek-chat',
  apiKey: process.env.DEEPSEEK_API_KEY,
});
// Excellent cost-performance for log parsing (~$0.0001 per call)

Anthropic (via compatible adapter)

const llm = new OpenAICompatibleProvider({
  provider: 'anthropic',
  model: 'claude-3-5-haiku-latest',
  apiKey: process.env.ANTHROPIC_API_KEY,
});

Custom Endpoint (self-hosted, proxy, or any compatible API)

const llm = new OpenAICompatibleProvider({
  provider: 'custom',
  model: 'my-fine-tuned-model',
  baseURL: 'https://my-inference-endpoint.example.com/v1',
  apiKey: 'sk-...',
});

Using the provider standalone (without pipeline)

import { OpenAICompatibleProvider } from '@agentix-e/log-parser-llm';

const llm = new OpenAICompatibleProvider({
  provider: 'deepseek',
  model: 'deepseek-chat',
  apiKey: process.env.DEEPSEEK_API_KEY,
});

const result = await llm.extractTemplate([
  'Disk usage on /dev/sda1 is 85% (threshold: 80%)',
  'Disk usage on /dev/sdb1 is 92% (threshold: 90%)',
]);

console.log(result.template);   // "Disk usage on <PATH> is <NUM>% (threshold: <NUM>%)"
console.log(result.variables);  // NER-style annotations with positions and categories
console.log(result.confidence); // 0.95
console.log(result.tokens);     // { input: ..., output: ... }

Configuration

OpenAICompatibleConfig

| Parameter | Type | Required | Default | Description | |-----------|------|:---:|---------|-------------| | provider | 'ollama' \| 'openai' \| 'deepseek' \| 'anthropic' \| 'custom' | Yes | -- | Backend provider type | | model | string | Yes | -- | Model identifier (e.g., 'qwen2.5:7b', 'gpt-4o-mini') | | apiKey | string | No | -- | API key. Not needed for Ollama. Can also be set via env var. | | baseURL | string | No | Provider default | Override the API endpoint URL |

Provider defaults

| Provider | baseURL fallback | |----------|--------------------| | ollama | http://localhost:11434/v1 | | openai | https://api.openai.com/v1 | | deepseek | https://api.deepseek.com/v1 | | anthropic | https://api.anthropic.com/v1 | | custom | http://localhost:11434/v1 |

How It Works

The provider uses the Vercel AI SDK's generateObject() with a Zod schema to enforce structured output. The LLM receives NER-style prompts identifying variable positions (IP, number, path, UUID, etc.) and replacing them with typed placeholders. Every response is validated against the schema before being returned to the pipeline.

The control plane flow: MissAccumulator buffers drain-ts misses → PartitioningEngine clusters by TF-IDF similarity → DppSampler selects representative samples → extractTemplate() sends NER prompt → SelfReflectionLoop validates output (up to 3 iterations) → AdaptiveTemplateCache stores the result.

API Reference

OpenAICompatibleProvider

| Method | Description | |--------|-------------| | new OpenAICompatibleProvider(config) | Create a provider. No network calls. | | extractTemplate(logSamples: readonly string[]) | Call the LLM with NER-style prompting. Returns LlmTemplateResult with template, variable annotations, confidence, and token usage. |

LlmTemplateResult

| Field | Type | Description | |-------|------|-------------| | template | string | Extracted template with typed placeholders (e.g., <IP>, <NUM>) | | variables | VariableAnnotation[] | NER-style annotations with position, value, and category | | confidence | number | Confidence score (0-1) | | tokens | { input: number; output: number } | Token usage for cost tracking |

Related Packages

| Package | Purpose | |---------|---------| | @agentix-e/log-parser-core | Core pipeline engine (required) | | @agentix-e/log-parser-webllm | Browser-local LLM via WebGPU (alternative to cloud providers) | | @agentix-e/log-parser-node | Node.js I/O adapters for file streaming | | @agentix-e/log-parser-server | REST API server wrapping the pipeline |

License

MIT