Datasheet Intelligence

Neural Inverse extracts structured hardware data from PDF datasheets using your configured LLM.

How It Works

1. Upload a PDF datasheet (reference manual, datasheet, or errata document)
2. Parse — the system extracts register maps, timing constraints, and errata
3. Inject — extracted data is added to the session context for AI completions
4. Cite — all generated code references datasheet page numbers

Uploading Datasheets

Via Firmware.inverse

{
  "datasheets": [
    "docs/stm32f407_reference_manual.pdf",
    "docs/stm32f407_datasheet.pdf"
  ]
}

Datasheets are parsed automatically at session start.

Via Agent Tool

> fw_upload_datasheet docs/peripheral_datasheet.pdf

What Gets Extracted

Register Maps

• Peripheral names and base addresses
• Register offsets, sizes, access types
• Bit field definitions with descriptions
• Reset values
• Page references for each register

Timing Constraints

• Setup and hold times
• Clock frequency limits
• Propagation delays
• Peripheral-specific timing (e.g., ADC sampling time, Flash wait states)
• Page references

Silicon Errata

• Erratum ID
• Affected peripheral
• Description and conditions
• Workaround
• Severity (info, minor, major, critical)
• Affected revisions
• Page reference

Querying

> fw_query_datasheet "USART baud rate calculation formula"

Searches extracted content and returns relevant information with page citations.

> fw_get_datasheet_citations "SPI clock polarity"

Returns specific page references where a topic is documented.

Code Generation with Citations

When the agent generates code with datasheet context active, it includes citations:

// Configure USART1 baud rate (RM0090 §30.3.4, p.978)
// BRR = fPCLK2 / (16 * baudrate) = 84MHz / (16 * 115200) = 45.57 ≈ 46
USART1->BRR = 46;

Storage

Extracted data is held in memory for the session duration. It's re-extracted each time the session starts (ensures freshness if datasheets are updated).

Agent Tools