Strengthen pintable layout_rules skill and refresh empty extractions.

Treat PCB/application layout guidance as a first-class extract, keep those pages in PDF trim, and re-run IC extraction once when rules are still empty under older model_version.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
2026-09-13 18:01:15 +02:00
co-authored by Cursor
parent 86121e6547
commit 21e2be1cd0
10 changed files with 332 additions and 64 deletions
+33
View File
@@ -4,6 +4,8 @@ from __future__ import annotations
from typing import Any
from packaging.version import Version
KNOWN_KINDS = frozenset({"decoupling_proximity", "thermal_via", "keepout", "length_match"})
@@ -20,6 +22,37 @@ def _num(v: Any) -> float | None:
return None
def has_any_layout_rule(raw: object) -> bool:
"""True when extraction already produced at least one structured rule."""
if not isinstance(raw, list):
return False
for row in raw:
if isinstance(row, dict) and str(row.get("kind") or "").strip() in KNOWN_KINDS:
return True
return False
def needs_layout_rules_refresh(
data: dict,
*,
min_scan_version: str,
) -> bool:
"""True when layout_rules are empty and the extract predates the scan version.
After a successful extract at ``min_scan_version`` or newer, an empty
``layout_rules`` list means the datasheet had no guidance — do not loop.
"""
if has_any_layout_rule(data.get("layout_rules")):
return False
ver = str(data.get("model_version") or "0.0.0")
if not min_scan_version or min_scan_version == "0.0.0":
return False
try:
return Version(ver) < Version(min_scan_version)
except Exception:
return True
def validate_layout_rules(raw: list | None) -> tuple[list[dict], list[str]]:
"""Return (normalized rows, errors). Empty list is a valid skip."""
if not raw:
+43 -4
View File
@@ -135,8 +135,36 @@ PINTABLE_TOOL = {
},
"layout_rules": {
"type": "array",
"description": "Optional PCB layout constraints from typical-application pages. kind must be decoupling_proximity, thermal_via, keepout, or length_match. max_distance_mm only if the PDF states a number — never invent 3 mm or 3W.",
"items": {"type": "object"},
"description": (
"PCB layout constraints from typical-application / PCB layout pages. "
"kind: decoupling_proximity | thermal_via | keepout | length_match. "
"Fields: pin, cap_value_hint, max_distance_mm (ONLY if the PDF states a "
"number — never invent 3 mm/JEDEC), same_layer (bool), min_via_count, "
"net_class, note, source_page. Empty array if the PDF has no layout guidance."
),
"items": {
"type": "object",
"properties": {
"kind": {
"type": "string",
"enum": [
"decoupling_proximity",
"thermal_via",
"keepout",
"length_match",
],
},
"pin": {"type": ["string", "null"]},
"cap_value_hint": {"type": ["string", "null"]},
"max_distance_mm": {"type": ["number", "null"]},
"same_layer": {"type": ["boolean", "null"]},
"min_via_count": {"type": ["integer", "null"]},
"net_class": {"type": ["string", "null"]},
"note": {"type": ["string", "null"]},
"source_page": {"type": ["integer", "null"]},
},
"required": ["kind"],
},
},
},
"required": ["component_subtype", "component_subtype_description", "package_info", "pintable"],
@@ -257,11 +285,17 @@ _MAX_PDF_PAGES = 120
log = logging.getLogger(__name__)
# Keywords used to find relevant pages for each extraction stage.
# Include PCB / typical-application pages so layout_rules can be extracted
# when large datasheets are trimmed to ≤_MAX_PDF_PAGES.
_PINTABLE_KEYWORDS = re.compile(
r"pin\s*(out|diagram|configuration|description|assignment|function|name|table|map)"
r"|ball\s*map|package\s*(pin|drawing|outline)|signal\s+description"
r"|absolute\s+maximum|recommended\s+operating|electrical\s+characteristics"
r"|ordering\s+information|device\s+information",
r"|ordering\s+information|device\s+information"
r"|pcb\s+layout|layout\s+(guideline|recommendation|consideration|hint)"
r"|typical\s+application|application\s+(circuit|schematic|information|note)"
r"|reference\s+design|decoupling|bypass\s+capacitor|thermal\s+via"
r"|land\s+pattern|keep[\s\-]?out|place\s+(close|near|within)",
re.IGNORECASE,
)
@@ -596,7 +630,12 @@ async def extract_pintable(
skill_name="extract-pintable",
model=model,
system=system,
user_text=f"Extract pin table and package info for MPN: {mpn}",
user_text=(
f"Extract pin table, package info, absolute maximum ratings, "
f"and layout_rules (scan PCB layout / typical application / "
f"thermal pages; max_distance_mm only if the PDF states mm) "
f"for MPN: {mpn}"
),
pdf_path=trimmed,
output_tool=_to_tool(PINTABLE_TOOL),
)
+19 -1
View File
@@ -762,6 +762,10 @@ async def _stage_ic_extraction(ctx: PipelineContext) -> None:
extracted_dir = ctx.ws.local_path("extracted")
# Pre-categorize: workspace cache, library cache, or needs extraction
from backend.config import settings as app_settings
from backend.pinscopex.layout_rules import needs_layout_rules_refresh
layout_scan_ver = app_settings.get_default_model_version()
_ic_cache: dict[str, tuple] = {}
_ic_new_count = 0
for mpn in ctx.ic_mpns:
@@ -771,11 +775,25 @@ async def _stage_ic_extraction(ctx: PipelineContext) -> None:
existing = json.loads(json_path.read_text())
if existing.get("pintable"):
ws_ver = existing.get("model_version", "0.0.0")
if not settings_svc.version_is_stale(ws_ver, ctx.min_ver):
if (
not settings_svc.version_is_stale(ws_ver, ctx.min_ver)
and not needs_layout_rules_refresh(
existing, min_scan_version=layout_scan_ver,
)
):
_ic_cache[mpn] = ("workspace",)
continue
lib_key = proj_svc.library_has_extraction(ctx.storage, mpn, min_version=ctx.min_ver)
if lib_key:
# Library hit may still lack layout_rules under the current skill.
try:
lib_payload = ctx.storage.read_json(lib_key)
except Exception:
lib_payload = {}
if not needs_layout_rules_refresh(
lib_payload if isinstance(lib_payload, dict) else {},
min_scan_version=layout_scan_ver,
):
_ic_cache[mpn] = ("library", lib_key)
continue
_ic_new_count += 1
+1 -1
View File
@@ -1,5 +1,5 @@
{
"default_model_version": "1.9.0",
"default_model_version": "1.10.0",
"extract-pintable": {
"skill_id": "skill_01VMWPZuvuZAe4LmLbmsNWNY",
"latest_version": "1784798970179642",
+2 -1
View File
@@ -25,8 +25,9 @@ Fonte originale: canvas *Pinscope: crescita e DeepSeek*. Qui lo stato operativo.
| Layout F1 | Domini / gruppi / satelliti | **Done**`functional_groups.json` (no mm) |
| Layout F1b | Pipeline Placement parallela | **Done** — API/UI `placement_*`, `placement_plan.json` |
| Layout F2 | Placement IC packing mm | **Partial** — skeleton gated (`placement_pack.json`); no zone/export yet |
| C4 | `layout_rules` dal datasheet | **Improved** — skill 1.10.0 + trim keywords + cache refresh se vuoto |
**Done when (prossimo pacchetto):** smoke `--live` verde; hit_ratio visibile in UI logs; F2 packing oltre skeleton (collisioni, zone, export).
**Done when (prossimo pacchetto):** smoke `--live` verde; hit_ratio visibile in UI logs; F2 packing oltre skeleton (collisioni, zone, export); più IC library con `layout_rules` numerici.
---
+9
View File
@@ -2,6 +2,15 @@
What's new in Pinscope.
## 2.28.8 — 2026-09-13 — Stronger layout_rules extraction skill
Pintable skill now treats PCB / typical-application layout guidance as a first-class extract. Page trim keeps layout keywords; IC cache re-extracts once when `layout_rules` are empty under an older `model_version` (1.10.0+).
- [Changed] `skills/extract-pintable` — dedicated `layout_rules` step, examples, hard negatives; tighter `validate.py`.
- [Changed] Datasheet page trim includes PCB layout / decoupling / typical-application keywords.
- [Changed] Analysis IC extraction refreshes empties when `model_version` < default (1.10.0).
- [Changed] `default_model_version``1.10.0`.
## 2.28.7 — 2026-09-13 — Placement F2 pack skeleton (gated)
Placement pipeline can propose satellite xy only when a `.kicad_pcb` layout exists and a `decoupling_proximity` rule has numeric `max_distance_mm`. No millimetres invented; otherwise `placement_pack.json` is skipped with an explicit reason.
+124 -55
View File
@@ -1,94 +1,163 @@
---
skill_name: extract-pintable
description: Extract pin table, package info, and component subtype from an IC datasheet PDF. Returns structured data via the save_pintable tool.
description: Extract pin table, package info, absolute-maximum ratings, layout_rules, and component subtype from an IC datasheet PDF. Returns structured data via the save_pintable tool.
---
# Extract Pin Table & Variant Info
Extract the pin table and variant/ordering information from a component datasheet and return it as structured JSON via the `save_pintable` tool.
Extract structured data from an IC datasheet and return it via the `save_pintable` tool.
**Priority order:** (1) complete pin table for the MPN package, (2) `layout_rules` from PCB / typical-application pages, (3) package + abs-max + subtype.
## Steps
### 1. Read the datasheet PDF
The datasheet PDF is provided in the user message. Focus on these sections:
- **Pin configuration / pin assignment table** — This is the primary target. Look for tables listing pin number, pin name, description, and alternate functions.
- **Ordering information / part number decoder** — Decode what each segment of the MPN means (package, temperature grade, packing, output voltage, etc.)
- **Package information** — Pin count, package type (QFP, BGA, SOT-23, etc.)
Focus on these sections (figures count as evidence):
- **Pin configuration / pin assignment table** — primary target
- **Ordering information / part number decoder**
- **Package information**
- **PCB layout / layout guidelines / land pattern notes**
- **Typical application / reference design** (placement callouts near caps, vias, keepouts)
- **Absolute maximum ratings**
### 2. Extract the pin table
For every pin on the component, extract:
- `number` (int or str) — The pin number, or BGA ball coordinate like `"A3"`
- `name` (str) — The pin name exactly as printed in the datasheet (e.g., `"VDD"`, `"PA0/SPI0_CLK"`)
- `description` (str or null) — A brief description if the datasheet provides one
- `functions` (list[str] or null) — Alternate/multiplexed functions if the pin supports them
For every pin:
- `number` (int or str) — pin number, or BGA ball like `"A3"`
- `name` (str) — verbatim from the datasheet (e.g. `"VDD"`, `"PA0/SPI0_CLK"`)
- `description` (str or null)
- `functions` (list[str] or null) — alternate/mux functions
Rules for pin extraction:
- Include ALL pins — power, ground, NC, and signal pins
- Use pin names verbatim from the datasheet — do not rename or normalize
- For multiplexed pins, put the primary name in `name` and alternates in `functions`
Rules:
- Include ALL pins — power, ground, NC, exposed pad / EP
- Names verbatim — do not rename or normalize
- Multiplexed pins: primary in `name`, alternates in `functions`
- If the datasheet has per-package tables, use the package matching the MPN
- Off-by-one pin numbers break everything downstream — double-check
Optional extras (omit if the PDF does not show them):
- `internal_features.pullup_pins` / `esd_clamp_pins` / `analog_switch` from the **block diagram** only.
- `layout_rules` from **PCB layout / typical application** pages. `kind` is only `decoupling_proximity`, `thermal_via`, `keepout`, or `length_match` (intra-pair skew mm). Set `max_distance_mm` only when the document states a number — do not invent JEDEC or USB millimetres.
Rules for pin extraction:`
- If the datasheet has separate tables for different packages, extract for the package matching the MPN
- Pay careful attention to pin numbering — off-by-one errors here break everything downstream
- **Modules vs bare die (critical).** MPNs containing `WROOM`, `WROVER`, `MODULE`, `MOD-`, or `SIP` are *modules*. Extract the **module landing-pad table** (connector pins the schematic uses). Do **not** extract the SoC/QFN ball map from a nested chip chapter or a sibling chip-only PDF.
- Espressif WROOM: pin 1 is GND (often a group of GND pads). Pin 1 named `ANT`, `CHIP_PU`, or `XTAL_*` means you grabbed the bare ESP32 die table — that will mark every module GND as “antenna shorted” and is invalid.
**Modules vs bare die (critical).** MPNs containing `WROOM`, `WROVER`, `MODULE`, `MOD-`, or `SIP` are *modules*. Extract the **module landing-pad table** (schematic pins). Do **not** extract the SoC/QFN ball map from a nested chip chapter.
- Espressif WROOM: pin 1 is GND. Pin 1 named `ANT`, `CHIP_PU`, or `XTAL_*` means you grabbed the die table — invalid.
- Crystal, RF antenna, and flash on a WROOM module are **inside the can**; they must not appear as schematic pin numbers.
### 3. Extract package info
Optional extras (omit if absent):
- `internal_features.pullup_pins` / `esd_clamp_pins` / `analog_switch` from the **block diagram** only.
Decode the MPN and package details into a single `PackageInfo`:
- `base_family` (str) — The base part family (e.g., `"MSPM0G3507"` from `"MSPM0G3507SPTR"`)
- `package` (str) — Package name (e.g., `"LQFP-48"`, `"SOT-23-5"`)
- `pin_count` (int) — Number of pins
- `description` (str) — Human-readable decoding of the full MPN (e.g., `"MSPM0G3507, 48-pin LQFP, tape & reel"`)
### 3. Extract layout_rules (required scan — empty OK)
Look for an "Ordering Information" or "Device Information" table in the datasheet — most datasheets have one.
You **must** look for layout guidance. Emit `layout_rules` as a list. Use `[]` only after scanning layout / application / thermal pages and finding no placement guidance.
### 4. Extract absolute maximum ratings
#### Where to look
- Headings: “PCB Layout”, “Layout Guidelines”, “Layout Considerations”, “Board Layout”, “Land Pattern”
- “Typical Application”, “Application Circuit”, “Reference Design”
- Thermal / EP / exposed-pad via recommendations
- Callouts on application figures (“place CIN within 2 mm of VIN”)
Copy the **Absolute Maximum Ratings** table (not Recommended Operating Conditions). For each row that a reviewer would need to compare against the schematic rails:
#### Allowed `kind` (closed set)
| kind | Use when |
| --- | --- |
| `decoupling_proximity` | Bypass / decoupling / input / output cap near a supply or pin |
| `thermal_via` | Vias under exposed pad / thermal pad / EP |
| `keepout` | Keep foreign nets, digital return, or copper out of a region |
| `length_match` | Intra-pair skew / matched length limit in mm |
- `parameter` (str) — as printed (`VCC`, `VIN`, `I/O pin voltage`, `Storage temperature`, …)
- `min` / `max` (number or null) — numeric limit; omit the other side if the table only lists one
- `unit` (str) — `V`, `mA`, `°C`, …
- `source_page` (int) — 1-based datasheet page of that row
#### Fields
- `pin` — number or name as printed (`"5"`, `"VIN"`, `"VDD"`, `"EP"`)
- `cap_value_hint` — only if shown (`"100nF"`, `"10µF"`)
- `max_distance_mm`**number only if the PDF states millimetres**
- OK: “within 2 mm”, “< 5 mm”, “no more than 3 mm from the pin” → `2` / `5` / `3`
- NOT OK as a number: “as close as possible”, “close to the pin”, “adjacent”, “nearby” → set `max_distance_mm: null` and keep the rule with a `note`
- **Never invent** JEDEC, USB, IPC, or “standard 3 mm / 5 mm” distances
- `same_layer``true`/`false` only if text says same side / opposite side of the board; else null
- `min_via_count` — integer only if stated (“at least 4 vias”)
- `net_class` / `note` — short quote of the guidance
- `source_page` — 1-based page of the guidance (required when you emit a rule)
Include supply voltages, pin/input voltages, input current, and temperature. Skip ESD *human-body-model / IEC contact-discharge kV* rows unless they are the only voltage limit given. Do not invent numbers; if the table is a raster with no readable values, return an empty array.
#### Examples
**ESD / TVS / protection ICs (`ic.protection.esd` and similar):** also copy from Electrical Characteristics (not only abs-max):
Numeric proximity (copy the millimetre from the PDF):
- Working / reverse working voltage **Vrwm** (or V_RWM / "operating voltage") as a **signed** min/max in volts — e.g. bidirectional ±13 V is `min: -13`, `max: 13`, `unit: "V"`.
- One extra row whose `parameter` states polarity/topology as printed (`bidirectional`, `unidirectional`, `back-to-back`), `unit: "—"`, min/max omitted. Do **not** infer unidirectional from "IO" pins that list GND as the reference pin.
```json
{
"kind": "decoupling_proximity",
"pin": "VIN",
"cap_value_hint": "10uF",
"max_distance_mm": 2.0,
"same_layer": true,
"note": "Place CIN within 2 mm of VIN",
"source_page": 14
}
```
### 5. Assign component subtype (taxonomy)
Proximity without a millimetre (still emit the rule):
The existing IC taxonomy subtypes are provided in the system prompt under `EXISTING IC TAXONOMY SUBTYPES`. Pick the best matching subtype based on the component's MPN, package info, and pin names.
```json
{
"kind": "decoupling_proximity",
"pin": "VDD",
"cap_value_hint": "100nF",
"max_distance_mm": null,
"note": "Place decoupling capacitor as close as possible to VDD",
"source_page": 22
}
```
If no existing subtype fits, propose a new one following the dot-notation convention (`ic.{category}.{specific}`).
Thermal vias:
Set the chosen subtype on the `component_subtype` field.
```json
{
"kind": "thermal_via",
"pin": "EP",
"min_via_count": 4,
"note": "Use at least 4 thermal vias in the exposed pad",
"source_page": 18
}
```
### 6. Quality checks
#### Hard negatives
- Do not invent land-pattern pad sizes from the mechanical drawing alone
- Do not emit `length_match` for USB/HDMI/PCIe unless the **this** datasheet states a skew/length number
- Do not use kinds outside the closed set
- One rule per distinct pin/guidance; prefer supply pins that show caps in the application figure
Before producing output, verify:
- Pin count matches what the datasheet says for this package
- No duplicate pin numbers
- No pins are missing (compare against the datasheet's stated pin count)
- Pin names look reasonable (not garbled OCR artifacts)
### 4. Extract package info
### 7. Validate and output
- `base_family` — e.g. `"MSPM0G3507"` from `"MSPM0G3507SPTR"`
- `package` — e.g. `"LQFP-48"`, `"SOT-23-5"`
- `pin_count` (int)
- `description` — human-readable MPN decode
Validate your extraction against the output schema:
Prefer “Ordering Information” / “Device Information” tables.
### 5. Extract absolute maximum ratings
Copy the **Absolute Maximum Ratings** table (not Recommended Operating Conditions):
- `parameter`, `min` / `max`, `unit`, `source_page` (1-based)
Include supply voltages, pin/input voltages, input current, temperature. Skip HBM/IEC kV ESD rows unless they are the only voltage limit. Do not invent numbers.
**ESD / TVS (`ic.protection.esd` and similar):** also from Electrical Characteristics:
- Vrwm / operating voltage as signed min/max in volts
- One row for polarity/topology as printed (`bidirectional`, …), `unit: "—"`
### 6. Assign component subtype
Pick the best dotted subtype from `EXISTING IC TAXONOMY SUBTYPES` (e.g. `ic.mcu`, `ic.power.ldo`). If none fit, propose `ic.{category}.{specific}`.
### 7. Quality checks
Before output:
- Pin count matches the package for this MPN
- No duplicate / missing pin numbers
- `layout_rules` scanned (list present; `[]` only if truly no guidance)
- Every emitted rule has a valid `kind`; every numeric `max_distance_mm` comes from the PDF text/figure
- Pin names are not OCR garbage
### 8. Validate and output
```bash
python3 /skills/extract-pintable/validate.py '<your JSON here>'
```
If validation passes, call the `save_pintable` tool with the structured result.
Do NOT write files to disk — use the tool.
If validation passes, call `save_pintable`. Do NOT write files to disk — use the tool.
+24 -1
View File
@@ -56,7 +56,30 @@
},
"layout_rules": {
"type": "array",
"items": {"type": "object"}
"description": "PCB layout constraints from typical-application / PCB layout pages. Empty if none stated.",
"items": {
"type": "object",
"properties": {
"kind": {
"type": "string",
"enum": [
"decoupling_proximity",
"thermal_via",
"keepout",
"length_match"
]
},
"pin": {"type": ["string", "null"]},
"cap_value_hint": {"type": ["string", "null"]},
"max_distance_mm": {"type": ["number", "null"]},
"same_layer": {"type": ["boolean", "null"]},
"min_via_count": {"type": ["integer", "null"]},
"net_class": {"type": ["string", "null"]},
"note": {"type": ["string", "null"]},
"source_page": {"type": ["integer", "null"]}
},
"required": ["kind"]
}
}
},
"required": ["component_subtype", "package_info", "pintable"]
+56
View File
@@ -93,6 +93,62 @@ def validate(data: dict) -> list[str]:
kind = row.get("kind")
if kind not in kinds:
errors.append(f"layout_rules[{i}] unknown kind: {kind!r}")
continue
dist = row.get("max_distance_mm")
if dist is not None and dist is not False:
if isinstance(dist, bool):
errors.append(
f"layout_rules[{i}].max_distance_mm must be a number or null"
)
elif isinstance(dist, (int, float)):
if float(dist) <= 0:
errors.append(
f"layout_rules[{i}].max_distance_mm must be > 0"
)
else:
try:
v = float(str(dist).strip())
except (TypeError, ValueError):
errors.append(
f"layout_rules[{i}].max_distance_mm must be numeric "
f"or null (got {dist!r}) — do not invent distances; "
f"use null when the PDF only says 'close'"
)
else:
if v <= 0:
errors.append(
f"layout_rules[{i}].max_distance_mm must be > 0"
)
via = row.get("min_via_count")
if via is not None and via is not False and not isinstance(via, bool):
if isinstance(via, int):
if via <= 0:
errors.append(
f"layout_rules[{i}].min_via_count must be > 0"
)
else:
try:
iv = int(float(str(via).strip()))
except (TypeError, ValueError):
errors.append(
f"layout_rules[{i}].min_via_count must be an integer "
f"or null (got {via!r})"
)
else:
if iv <= 0:
errors.append(
f"layout_rules[{i}].min_via_count must be > 0"
)
page = row.get("source_page")
if page is not None and not isinstance(page, int):
errors.append(
f"layout_rules[{i}].source_page must be an integer or null"
)
same = row.get("same_layer")
if same is not None and not isinstance(same, bool):
errors.append(
f"layout_rules[{i}].same_layer must be a boolean or null"
)
return errors
+20
View File
@@ -33,6 +33,26 @@ def test_empty_list_is_explicit_skip():
assert errors == []
def test_needs_refresh_when_empty_and_old_version():
from backend.pinscopex.layout_rules import needs_layout_rules_refresh
assert needs_layout_rules_refresh(
{"model_version": "1.9.0", "layout_rules": []},
min_scan_version="1.10.0",
)
assert not needs_layout_rules_refresh(
{"model_version": "1.10.0", "layout_rules": []},
min_scan_version="1.10.0",
)
assert not needs_layout_rules_refresh(
{
"model_version": "1.9.0",
"layout_rules": [{"kind": "decoupling_proximity", "max_distance_mm": None}],
},
min_scan_version="1.10.0",
)
def test_unknown_kind_rejected_and_non_numeric_distance_is_null():
ok, errors = validate_layout_rules([
{"kind": "not_a_kind", "max_distance_mm": 1.0},