Datasheets Skill

Related Skills

Handoff guidance: This skill is consumer infrastructure. The typical flow is distributor skill downloads PDF → datasheets skill extracts → analyzer skill queries. Use this skill directly when (a) the user asks to extract or verify a specific MPN, (b) an analyzer reports trust_level: low and the gap is per-MPN extraction quality, or (c) a new MPN was added to the BOM and downstream detectors should pick up its verified specs. Don't run this skill in isolation if the user just wants a design review — call it from the kicad workflow at the "Sync datasheets" step instead.

Purpose

Extract structured, machine-readable specifications from component datasheet PDFs and make them available to analyzer skills. Works on whatever PDFs are downloaded under <project>/datasheets/ (downloads are owned by distributor skills like digikey, mouser, lcsc, element14).

Scope

This skill owns:

• Extraction schemas — canonical JSON structures for per-MPN specs. v1.4 ships 6 JSON Schema Draft 2020-12 schemas under schemas/ (base, pinout, specvalue, regulator, extraction, manifest) plus 5 v1.4 category extensions (diode, transistor, opamp, mcu, crystal). v1.3 cache format (EXTRACTIONVERSION in scripts/datasheetextractcache.py) is still read for compat.
• Typed access layer (v1.4) — datasheettypes/ package exposes DatasheetFacts, SpecValue, Pin, Pinout, lookup(), best(), trusted(), hasdata(). Recommended for all new consumers.
• PDF page selection — heuristics to pick pages most likely to contain pinouts, e-chars, applications, SPICE models.
• Quality scoring — v1.4 uses a three-dimension rubric (pinout completeness, base completeness, category-extension completeness, 0–100 scale). v1.3 5-dimension weighted rubric still applies to legacy caches.
• Consumer APIs — scripts/datasheetlookup.py for v1.4 typed access; scripts/datasheetfeatures.py for the v1.3 dict-shaped helpers (getregulatorfeatures, getmcufeatures, getpinfunction) — the v1.3 helpers dual-read v1.4 caches and translate to v1.3 dict shape for legacy detector code. Sunset planned for v1.6.
• Verification — datasheetverify.py (v1.3, schema-vs-usage cross-check) plus datasheetverifyv14extraction (v1.4, power_domain references resolve, recommended ≤ absolute, regulator pin references exist).

Non-goals

• No PDF downloading. That is owned by distributor skills (digikey, mouser, lcsc, element14).
• No global library. Each project's extractions live in <project>/datasheets/extracted/. There is no shared cross-project cache.

Cache location

<project>/
  design.kicad_sch
  datasheets/
    TPS61023DRLR.pdf        # downloaded by distributor skills
    extracted/
      manifest.json         # extraction manifest (legacy name: index.json)
      TPS61023DRLR.json     # structured extraction (this skill's output)

Reference guides

• references/extraction-schema.md — canonical schema, every field defined
• references/field-extraction-guide.md — how to find each field in datasheets from common vendors (TI, ST, NXP, Espressif, Microchip)
• references/quality-scoring.md — rubric details, score thresholds
• references/consumer-api.md — how kicad/emc/spice/thermal consume extractions
• references/cache-layout.md — v1.4 cache directory convention (per-MPN files, _families/ reservation, staleness rules)

Entry-point scripts

• scripts/datasheetextractcache.py — v1.3 cache manager, resolver, indexer
• scripts/datasheetpageselector.py — page selection heuristics (used by both v1.3 and v1.4 pipelines)
• scripts/datasheet_score.py — v1.3 extraction quality scoring
• scripts/datasheetverify.py — cross-check extraction vs schematic usage (v1.3 + v1.4 verifyv14_extraction mode)
• scripts/datasheet_lookup.py — v1.4 typed lookup(mpn) → DatasheetFacts facade with staleness detection
• scripts/datasheetfeatures.py — v1.3 consumer helper API (dual-reads v1.4 caches via derive*v14 translators)
• scripts/plan_extraction.py — v1.4 orchestration plan generator (Phase 3 extraction pipeline)
• scripts/merge_results.py — v1.4 per-task result validator + merger
• datasheettypes/ — v1.4 typed access layer package (DatasheetFacts, SpecValue, Pin, Pinout, lookup, best, trusted, hasdata)

Extraction workflow

Run python3 skills/datasheets/scripts/planextraction.py <project> to generate an orchestration plan, then mergeresults.py to validate and merge per-task outputs. Full scout→plan→dispatch→merge procedure: [references/extraction-pipeline.md](references/extraction-pipeline.md).

Consuming extractions (v1.4 typed API)

The recommended consumer surface is the typed lookup(mpn, cachedir=...) facade plus the trust-gating helpers from datasheettypes. Import like:

import sys, pathlib
sys.path.insert(0, str(pathlib.Path(__file__).parent.parent / "datasheets"))
from datasheet_types import lookup, has_data, best, trusted

# Returns Optional[DatasheetFacts]. None on cache miss / stale PDF / low quality.
facts = lookup("TPS61023DRLR", cache_dir=pathlib.Path("datasheets/extracted"))
if facts is None:
    return  # heuristic-only path; no datasheet evidence available

# Field-level trust gating — every SpecValue list runs through has_data() / best() / trusted().
pu_range = facts.base.recommended_pullup_range  # Optional[list[SpecValue]]
if has_data(pu_range):
    # Most-trusted single value (first SpecValue meeting threshold, preserves extractor order).
    rec = best(pu_range, min_confidence="medium")  # Optional[SpecValue]
    if rec is not None and rec.min is not None:
        ...  # use rec.min, rec.max, rec.typ, rec.unit, rec.evidence.{page,section,confidence}

# All SpecValues at threshold (for multi-value fields like absolute_max).
hi_conf = trusted(facts.base.absolute_max.get("VDD", []), min_confidence="high")

Defensive patterns (mirrors kicad/SKILL.md § "Probing Analyzer JSON"):

• lookup() returns None on cache miss, stale PDF (PDF newer than extraction), or quality score below the configured floor. Always guard with if facts is None: return.
• Category extensions are optional on DatasheetFacts. facts.regulator is None when the part isn't in the regulator category — check before dereferencing.
• SpecValue lists can be None (field not extracted), [] (extracted but empty), or list[SpecValue]. has_data() collapses the first two to False; pair with best() / trusted() for confidence gating.
• SpecValue.min / .max / .typ are each Optional[float]. A SpecValue carrying only typ (no range) makes > / < comparisons against .min / .max raise TypeError — guard with explicit is not None chains on every numeric access.
• confidence is one of "low" / "medium" / "high". Calling best() / trusted() with any other string raises ValueError.

v1.3 compat shim

Legacy detectors still call getregulatorfeatures(mpn) / getmcufeatures(mpn) / getpinfunction(mpn, pin) from scripts/datasheet_features.py. These dual-read v1.4 caches and translate to the v1.3 dict shape. Sunset planned for v1.6 — new code should use lookup() directly.

When to trigger this skill

• Immediately after downloading datasheets via syncdatasheetsdigikey.py, syncdatasheetslcsc.py, or equivalent. Without extraction, IC-aware checks (VM-001 rail voltage, PS-001 power-good, PR-004 USB, DP-002 USB speed classification) fall back to heuristics on unknown ICs.
• Before running analyzers on a new project where datasheets are present but datasheets/extracted/ is empty — the analyzers won't produce the extractions themselves.
• When a review flags low trust level due to missing manufacturer evidence: extracting the ICs referenced by power regulators, MCUs, and high-speed peripherals typically flips trust_level: low → mixed or high.
• When a user asks for pin verification ("verify U1 pin names match datasheet") — this skill's cached extraction is the authoritative source.