michele/periscope
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Do not stall Retry failed on re-extraction or serial datasheet lookups.

Browse Source 
Resume reuses pintables, skips LCSC/DigiKey before review, and emits a waiting heartbeat so a long DeepSeek turn is visible instead of a freeze.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
michele
2026-08-28 08:34:17 +02:00
co-authored by Cursor
parent a9238875dd
commit 66ae877b7b
5 changed files with 101 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/services/llm/deepseek_provider.py
+3
View File
@@ -17,6 +17,7 @@ import logging
from typing import Any
from openai import APIStatusError, AsyncOpenAI
import httpx
from backend.config import settings
from backend.services.llm.base import LLMProvider, LLMSession
@@ -358,6 +359,8 @@ class DeepSeekProvider(LLMProvider):
self._client = AsyncOpenAI(
api_key=api_key,
base_url=settings.deepseek_base_url,
timeout=httpx.Timeout(420.0, connect=20.0),
max_retries=1,
)
async def create_session(
backend/services/llm/pdf_ingest.py
+29 -1
View File
@@ -198,6 +198,34 @@ def pdf_to_openai_content(
max_images: int = _DEFAULT_MAX_IMAGES,
) -> list[dict]:
"""OpenAI-style content parts for one PDF: text, plus images if vision."""
pdf_path = Path(path)
try:
mtime = pdf_path.stat().st_mtime_ns
except OSError:
mtime = 0
key = (str(pdf_path.resolve()), mtime, vision, max_chars, max_images)
cached = _PDF_CONTENT_CACHE.get(key)
if cached is not None:
return cached
parts = _pdf_to_openai_content_uncached(
pdf_path, vision=vision, max_chars=max_chars, max_images=max_images,
)
if len(_PDF_CONTENT_CACHE) > 32:
_PDF_CONTENT_CACHE.clear()
_PDF_CONTENT_CACHE[key] = parts
return parts
_PDF_CONTENT_CACHE: dict[tuple, list[dict]] = {}
def _pdf_to_openai_content_uncached(
path: Path,
*,
vision: bool,
max_chars: int,
max_images: int,
) -> list[dict]:
text = extract_pdf_text(path, max_chars=max_chars)
parts: list[dict] = [{"type": "text", "text": text}]
if not vision:
@@ -209,7 +237,7 @@ def pdf_to_openai_content(
"type": "text",
"text": (
f"The following {len(images)} image(s) are rendered pages of "
f"{Path(path).name} (pin tables, abs-max, electrical, and "
f"{path.name} (pin tables, abs-max, electrical, and "
f"application sections preferred over the front matter). "
f"Use them for diagrams and tables that text extraction may have missed."
),
backend/services/pipeline.py
+62 -5
View File
@@ -397,6 +397,10 @@ class PipelineContext:
# API call is still made (and the USD cost is still recorded in logs),
# but nothing is charged to the user's balance.
free: bool = False
# Reprocess of failed reviews / credit resume: do not re-extract
# pintables (vision can stall for many minutes) and do not block the
# review stage on LCSC/DigiKey lookups for ICs that already missed.
resume: bool = False
# ---------------------------------------------------------------------------
@@ -725,8 +729,20 @@ async def _ensure_local_datasheet(
return True
def _prior_extract_ready(ctx: PipelineContext) -> bool:
graph = ctx.ws.local_path("design_graph.json")
extracted = ctx.ws.local_path("extracted")
return graph.is_file() and extracted.is_dir() and any(extracted.glob("*.json"))
async def _stage_ic_extraction(ctx: PipelineContext) -> None:
"""Stage 2 — Extract IC pin tables from datasheets."""
if ctx.resume and _prior_extract_ready(ctx):
broker.publish(ctx.project_id, "step_update",
{"stage": "ic_extraction", "status": "complete",
"detail": "reusing previous extraction"})
return
extracted_dir = ctx.ws.local_path("extracted")
# Pre-categorize: workspace cache, library cache, or needs extraction
@@ -862,6 +878,11 @@ async def _stage_simple_extraction(ctx: PipelineContext) -> None:
"""Stage 2.5 — Extract specs for discrete/simple components."""
if not ctx.simple_mpns:
return
if ctx.resume and _prior_extract_ready(ctx):
broker.publish(ctx.project_id, "step_update",
{"stage": "simple_extraction", "status": "complete",
"detail": "reusing previous extraction"})
return
models_dir = ctx.ws.local_path("models")
@@ -1220,6 +1241,12 @@ async def _stage_passive_extraction(ctx: PipelineContext) -> None:
ctx.patterns = load_patterns(str(patterns_dir)) if patterns_dir.is_dir() else []
if ctx.resume:
broker.publish(ctx.project_id, "step_update",
{"stage": "passive_extraction", "status": "complete",
"detail": "reusing previous extraction"})
return
unresolved: dict[str, list[str]] = {}
for mpn, refs in ctx.passive_mpns.items():
if resolve_mpn(mpn, ctx.patterns) is not None:
@@ -1491,11 +1518,11 @@ async def _stage_validation(ctx: PipelineContext) -> None:
derating_path = ctx.ws.local_path("derating.json")
derating_path.write_text(json.dumps(derating_rows, indent=2) + "\n")
# Ensure all IC datasheet PDFs are available locally for review.
# Ensure IC datasheet PDFs are available locally for review.
# Cached ICs skipped pintable extraction, so their PDFs may not
# have been downloaded yet. Auto-fetch fills remaining gaps.
# Ensure PDFs for BOM ICs and any extra U* the netlist classified as IC.
seen_mpn: set[str] = set()
# On resume, skip the network lookup — LCSC/DigiKey for ICs that
# already missed can stall the review stage for many minutes.
mpns_to_place = list(ctx.ic_mpns)
for comp in ctx.graph.components.values():
if comp.component_type != ComponentType.IC:
@@ -1503,14 +1530,43 @@ async def _stage_validation(ctx: PipelineContext) -> None:
extra = (comp.mpn or "").strip()
if extra:
mpns_to_place.append(extra)
unique_mpns: list[str] = []
seen_mpn: set[str] = set()
for mpn in mpns_to_place:
if mpn in seen_mpn:
continue
seen_mpn.add(mpn)
unique_mpns.append(mpn)
async def _place(mpn: str) -> None:
safe = safe_mpn(mpn)
pdf_path = ds_dir / f"{safe}.pdf"
if not pdf_path.is_file():
await _ensure_local_datasheet(ctx, mpn, pdf_path, stage="review")
if pdf_path.is_file():
return
if ctx.resume:
from backend.services.datasheet_finder import find_local_pdf, mpn_query_variants
alt = find_local_pdf(ds_dir, mpn)
if alt is not None and alt.is_file() and alt.resolve() != pdf_path.resolve():
pdf_path.write_bytes(alt.read_bytes())
return
for name in mpn_query_variants(mpn) or [mpn]:
lib_ds_key = proj_svc.library_has_datasheet(ctx.storage, name)
if lib_ds_key:
ctx.storage.download_to_local(lib_ds_key, pdf_path)
return
return
try:
await asyncio.wait_for(
_ensure_local_datasheet(ctx, mpn, pdf_path, stage="review"),
timeout=45,
)
except TimeoutError:
logger.warning("Datasheet lookup timed out for %s; reviewing without it", mpn)
except Exception:
logger.exception("Datasheet lookup failed for %s", mpn)
await asyncio.gather(*(_place(m) for m in unique_mpns))
# Snapshot the full review queue so pause checkpoints can show what's left.
# Mirrors the filter in validate_design_async: ICs with a PDF available.
@@ -1701,6 +1757,7 @@ async def run_pipeline(
meta=meta,
min_ver=min_ver,
free=free,
resume=resume,
)
# On resume: carry over prior per-IC review completion so
backend/services/validation.py
+6
View File
@@ -379,6 +379,12 @@ async def review_ic_async(
tools = _ALL_TOOL_SCHEMAS
tool_choice = "auto"
if on_progress:
await on_progress(
ic_ref, turn, "waiting",
f"model turn {turn + 1}/{_MAX_REVIEW_TURNS}",
)
completion = await session.complete(
messages=messages,
tools=tools,
frontend/src/hooks/use-pipeline-progress.ts
+1
View File
@@ -14,6 +14,7 @@ const CACHED_DETAILS = new Set([
"specs from library",
"already resolved",
"all passives already resolved",
"reusing previous extraction",
]);
// Single source of truth for stage order and display metadata.