RTFM

Robots That Find Manuals (RTFM)

A collection of AI agents that actively research, update, and maintain a deterministic MCP server exposing solar-equipment fault, diagnosis, and remediation knowledge.

Solar O&M needs an AI native knowledge repo

Deployed solar is growing faster than the workforce that keeps it running. Every fleet commissioned today is a future stream of fault codes, blinking LEDs, and service calls, and the expertise to handle them is scarce, expensive, and walks out the door when a senior tech moves on.

43 GW

of new US solar capacity added in 2025, solar's fifth straight year as the #1 source of new generating capacity.

SEIA / Wood Mackenzie

86%

of solar employers report difficulty hiring qualified workers, with insufficient skills and training the top cited obstacle.

IREC 2024 Solar Jobs Census

The knowledge needed to diagnose any fault already exists. It's just buried in thousands of manufacturer PDFs: inconsistent formats, scattered vendor portals, a different document for every model revision.

Ramping up expertise can't mean reading manuals faster. It needs a single, structured, machine-readable repository of equipment knowledge: AI-native from the ground up, ready to plug into whatever agent platform an individual installer or a national O&M fleet already runs.

RTFM is an AI agent–maintained knowledge base

This service synchronizes daily with the California Energy Commission's equipment-certification registry: the canonical list of every string inverter, micro-inverter, and solar battery certified for sale in the largest solar market in the US.

From there, AI agents do the work no human team could sustain. A research agent scours the web for official documentation for every registered model. A triage agent verifies each candidate document and matches it to the exact equipment it covers. An extraction pipeline reads every page, text and figures, and distills it into structured fault knowledge.

CEC registry sync

Daily sync with ca.gov's certified-equipment lists

Research agent

Searches the web for official documentation per model

Triage agent

Verifies each document and matches it to exact equipment

Analysis & extraction

Reads every page, text and figures, into fault knowledge

Structured data

Served deterministically over MCP

206

manufacturers

3,031

models tracked

94%

of registered models documented

1,325

manuals ingested

35,254

pages indexed

16,124

alert codes

2,154

fault behaviours

24/7

monitoring & updates

Live numbers, straight from the database.

Document coverage of the US market

Providing your AI agents with up-to-date and comprehensive knowledge of US solar tech

Different problems need different altitudes. The data and MCP tool surface are designed so an agent can drop to the exact page of one specific manual, or rise above any single manufacturer and reason about solar faults in general.

The library tier

A fully processed, indexed library of every ingested document. Each page is converted to searchable markdown plus imagery, so multimodal agents don't just read the documentation, they see it: wiring diagrams, LED state tables, connector photos. Every retrieval cites its exact source page.

The structured tier

Pre-extracted tables of the data that matters in the field: 16,124 alert codes and 2,154 fault behaviours, organized by manufacturer and model, each carrying meaning, root cause, and recommended actions, with a page reference back to the manual that says so.

The abstraction tier

Across 14,703 analyzed alerts, every fault decomposes into four axes: the component involved, the state it's in, what the device did automatically, and what a human is told to do next. A conditional belief model built on the co-occurrence of those axes lets agents generalize across all manufacturers, surfacing the most likely diagnoses and best-documented paths to resolution even when the exact manual isn't on file.

ComponentStateAutomated responseHuman action
How 14,703 documented faults flow from component → state → automated response → human action. Link width = co-occurrence frequency across all manufacturers.
The controlled vocabulary behind the belief model: 485 fault concepts across four axes. Segment size = prevalence in the corpus. Click a segment to drill down.