data-analyzer

Large-scale JSON/JSONL data analysis CLI using local LLMs.

Uses a sliding window + progressive summarization approach to overcome context window limitations — no Map-Reduce information loss.

Features

• Analyze up to 100K+ JSON/JSONL records with local LLMs
• Sliding window engine with overlap for boundary context preservation
• Citation verification — every finding cites source records, verified against originals
• Checkpoint-based resume for long-running analysis
• Idempotent job execution
• Interactive parameter builder (with file input support)
• Markdown and HTML report output
• Output language control (--lang Japanese)

Requirements

• Go 1.23+
• Local LLM with OpenAI-compatible API (e.g., LM Studio)
• Recommended model: google/gemma-4-26b-a4b (Think OFF)

Installation

make build    # → dist/data-analyzer

Setup

# Option 1: Environment variables
export DATA_ANALYZER_API_ENDPOINT="http://localhost:1234/v1"
export DATA_ANALYZER_API_MODEL="google/gemma-4-26b-a4b"

# Option 2: Config file (~/.config/data-analyzer/config.toml)
mkdir -p ~/.config/data-analyzer
cp config.example.toml ~/.config/data-analyzer/config.toml

Usage

1. Prepare analysis parameters

Build parameters interactively with LLM assistance:

# Interactive mode (supports multi-line input, end with empty line)
data-analyzer prepare --output params.json

# With sample data — LLM sees actual field names and values
data-analyzer prepare --sample logs.jsonl --output params.json

# Load requirements from file + sample data, then refine interactively
data-analyzer prepare --input requirements.txt --sample logs.jsonl --output params.json

Or create params.json manually:

{
  "perspective": "Detect insider threats and unauthorized access",
  "target_fields": ["user", "action", "source_ip", "timestamp"],
  "attention_points": [
    "Multiple failed login attempts",
    "Privilege escalation",
    "Large data transfers to external services"
  ],
  "user_findings": [],
  "lang": "Japanese"
}

2. Run analysis

# Single file
data-analyzer analyze --params params.json logs.jsonl

# Directory (all .json/.jsonl files)
data-analyzer analyze --params params.json ./log_dir/

# With output file and language
data-analyzer analyze --params params.json --lang Japanese --output result.json logs.jsonl

# Resume interrupted analysis
data-analyzer analyze --params params.json --resume <job-id> logs.jsonl

3. Generate report

# Markdown to stdout
data-analyzer compile result.json

# HTML report
data-analyzer compile --format html --output report.html result.json

# Both Markdown and HTML
data-analyzer compile --format both --output report result.json

# From stdin
cat result.json | data-analyzer compile -

Configuration

Settings are loaded in order: defaults → config file → env vars → CLI flags.

Variable	Default	Description
DATA_ANALYZER_API_ENDPOINT	http://localhost:1234/v1	OpenAI-compatible API endpoint
DATA_ANALYZER_API_MODEL	google/gemma-4-26b-a4b	Model name
DATA_ANALYZER_API_KEY	—	API key (optional)
DATA_ANALYZER_CONTEXT_LIMIT	131072	Context window budget (tokens)
DATA_ANALYZER_OVERLAP_RATIO	0.1	Window overlap ratio (0.0–1.0)
DATA_ANALYZER_MAX_FINDINGS	100	Max findings to accumulate
DATA_ANALYZER_MAX_RECORDS_PER_WINDOW	200	Max records per window (quality guard)
DATA_ANALYZER_LANG	—	Output language (e.g., Japanese)
DATA_ANALYZER_TEMP_DIR	$TMPDIR/data-analyzer	Checkpoint directory

Resilience

When the LLM backend crashes or unloads the model during long analysis sessions, the client automatically detects the failure and waits for the model to reload. Configurable via [resilience] in config.toml:

Setting	Default	Description
max_retries	10	Max retry attempts per LLM call
max_backoff_sec	120	Max backoff wait between retries (seconds)
health_check_interval_sec	10	Polling interval for model readiness (seconds)
health_check_timeout_sec	300	Max wait for model to become ready (seconds)

On each retry, the client polls /v1/models to confirm the model is loaded before sending the next request. This prevents wasting retries while the backend is still reloading.

How It Works

┌─────────────┐    ┌──────────────┐    ┌──────────────┐
│   prepare    │───▶│   analyze    │───▶│   compile    │
│ (interactive)│    │(sliding win) │    │(md/html/both)│
└─────────────┘    └──────────────┘    └──────────────┘
   params.json        result.json       report.md/.html

Sliding Window Algorithm:

1. Divide records into overlapping windows (max 200 records per window)
2. For each window: [Previous Summary] + [Findings] + [New Data] → LLM
3. LLM returns updated summary + new findings with record citations
4. Citation verification: check excerpt relevance, replace with full original record
5. Checkpoint saved after each window (resume on interruption)
6. Final report generated from accumulated findings

Citation Verification:

Every citation from the LLM is verified against the original data:

• Excerpt values checked for relevance against the original record
• Excerpts always replaced with the full original record (no field omission)
• Non-matching excerpts flagged as possible hallucination
• Missing citations recovered from Record #N references in description text

Memory Map (128K token budget):

Section	Allocation
System prompt	~2K (fixed)
Previous summary	0→15K (grows, then stabilizes)
Accumulated findings	0→20K (grows, priority eviction)
New RAW data	Remainder (~86K–106K)
Response buffer	~5K (fixed)

4. Clean up old jobs

# Remove completed jobs older than 7 days (default)
data-analyzer clean

# Remove completed jobs older than 1 day
data-analyzer clean --max-age 1d

# Remove all jobs (including incomplete)
data-analyzer clean --all

License

MIT