PulseForge

AI-driven pipeline that turns unstructured log streams into categorized, templated, health-scored signal — a sibling project to ParseForge, reusing its trial → integration → promotion template pipeline instead of reimplementing it.

Full design plan: SPEC.md.

Why

CLI-command output ("show interface status") and log-body messages ("%LINK-3-UPDOWN: Interface ... changed state to down") are both just "unstructured text that recurs in a small number of shapes and needs a template." ParseForge already solves the hard part of that problem — LLM-drafted templates, self-validated against their own sample, clustered by output schema, and promoted only once they clear a match-rate gate, with a human-reviewed path for anything short of it. PulseForge applies that same machinery to log messages, and adds the two things logs need that CLI output doesn't:

• Envelope splitting — a log line isn't one blob, it's a <date-time-marker>: <log-body> pair. The marker has to be split off before anything downstream (naming, templating) sees clean body text.
• Category naming instead of command naming — there's no fixed command vocabulary to resolve against; an AI has to look at a log body and name its type (LINK-3-UPDOWN) before a template can even be requested for it.
• Pulse (health) scoring — once a category has an authoritative template, PulseForge tracks its match-rate, frequency, and field-value distribution over time and flags when a category goes unhealthy or inconsistent — not just "the template broke" (drift), but "this category's behavior looks wrong."

Architecture

ingestion -> envelope split -> category naming -> template forge (delegates
to parseforge) -> parse -> pulse/health scoring -> sink

See SPEC.md for the full breakdown of each stage.

Status

Scaffold only. Directory layout and module boundaries are in place; pipeline stages are stubs (NotImplementedError) pending real implementation. Nothing here is wired end to end yet.

Installation

# local development
pip install -e ".[dev,sampling]"

Relationship to ParseForge

PulseForge depends on parseforge as an ordinary pip package — the template-forging stage (pulseforge/forge/adapter.py) calls straight into parseforge.api rather than forking or copying its pipeline. A PulseForge "category" is passed into that pipeline the same way ParseForge's own CLI resolves and passes a "cli-name": as the key a group of samples is generated, validated, and promoted under.