The accept='' attribute and help copy already allowed .srt, but a
separate ALLOWED_EXTENSIONS array in upload.html's JS was filtering
out .srt files as 'unsupported format'. Adding 'srt' to that array
fixes the silent skip seen on Dev (file picker showed .srt as
valid, then the submit handler dropped it).
Profile YAML is descriptive metadata (executor runs unconditionally).
Documenting srt_structure (15), srt_timing (10), srt_language (20)
so the profile page reflects the live check set.
Upload form accepts .srt alongside .mp4. Configure page shows pair_map
counts and a collapsible list of unpaired SRTs (rendered via DOM
textContent to avoid XSS from user-controlled SRT filenames). Uses the
new /pairing-preview/<session_id> endpoint.
Adds optional srt_paths constructor parameter. At execute() top, runs
pair_batch() to produce pair_map / unpaired_srts / unpaired_videos.
Threads pair_map[video_path] into each per-video VideoQCExecutor as
srt_path. No-op when srt_paths is empty.
Skipped when self.srt_path is None (one result per check, weight set so
the weighted-average math is unchanged). When set, runs all three
checks sequentially with progress updates. SRT results appear as
additional cards in the existing Video QC report.
Text-only LLM call samples up to 15 cues (~1500 chars), asks Gemini
Flash to identify language. Pass = ISO matches expected from video
filename's locale; warning = low confidence or mixed_language; fail =
ISO mismatch with high confidence. Weight 20.
Note: uses genai.GenerativeModel directly rather than a unified
LLMConfig.call_text_api (which doesn't exist yet). Marked TODO for
future refactor when that helper is added.
score_pair: additive locale (0.5) + campaign code (0.3) + clip-slug
substring (0.4), capped at 1.0, with hard-reject on divergent locales
or non-overlapping slugs. pair_batch: greedy highest-first assignment
above 0.7 threshold; one SRT per video.
Verified pairs all 6 videos in testing_15may/srt/ to their SRTs.
Extract campaign code, clip slug, and locale from both video and SRT
filenames. Handles the two SRT styles seen in testing_15may/srt/
(campaign-code-prefixed CFUL... form, and abbreviated RIO_INTRO6B form).
Verified at REPL against test data.
Pure-function helpers, verified at REPL. canonical_locale handles the
de-AT ↔ AT-de order flip between SRT and video filenames; normalise_slug
strips non-alphanumerics so RIO_INTRO_15C ≈ RIO_INTRO15C.
Profile YAML is descriptive metadata (executor runs unconditionally).
Keeping it current so the profile page and any future YAML-driven
selection reflects the live check set.
Flags (never fails) when price or garment-name text falls inside known
platform UI overlay zones (TikTok / IG Stories / IG Reels / generic
vertical). Platform inferred from filename tokens via _infer_platform_zones.
Weight 0 in profile — advisory only, never contributes to overall score.
Single Gemini direct-video call detects garment/product text overlays;
deterministic match against PricingReference.get_prices() product_name
for the file's locale. Skips when no pricing reference attached, locale
unparseable, GEN/CEN file, no expected product names for locale, or no
on-screen garment text detected. Weight 25 in standard_video profile.
Adds _normalize_product_name (lowercase, alphanumeric+space, collapse
whitespace) and _product_names_match (substring or >=60% token-set
overlap on min side). Used by the upcoming garment_name check.
Adds _PLATFORM_ZONES (TikTok / IG Stories / IG Reels / generic vertical)
and _infer_platform_zones(filename) for use by the new title_safe check.
Pure function, verified at REPL against expected filenames. No new
behaviour exposed yet — wired up in the next task.
The price_currency check has always done a full numeric match against
the pricing reference but the report card only showed pass/fail by
currency. Pull matched_price, matched_product, detected_prices, and
expected_prices into the message string so QC reviewers can see the
full match at a glance.
No logic changes.
Single-file QC populated executor.context['user'] from current_user_email()
in routes.py, but batch QC routed through BatchQCExecutor — which never
accepted a user kwarg or set context['user'] on its per-file QCExecutor
instances. Result: every LLM call from a batched HM QC run logged as
anonymous in the Usage dashboard, only single-file and Video QC runs
showed the user's email.
BatchQCExecutor now takes user and stamps it onto each per-file
executor's context just before execute(), matching the Video QC
batch executor pattern.
Mirrors the hint pattern just added to Video Master so users can see
exactly which Box folder the search is scanning, with a clickable
link to open it in Box for self-diagnosis when a job number doesn't
turn up.
The earlier swap to BOX_CAMPAIGNS_FOLDER_ID=133295752718 was wrong —
Video Master operates on the automation campaigns folder
(156182880490), where subfolders are named by campaign TITLE rather
than the numeric job ID used in Reporting's root.
Reverted the default in config.py and all three .env example files.
Folder naming on Box is inconsistent — '1_CFUL263C01C_Kids drop1' vs
'1_CFUL263C01F-Kids drop 2' vs 'Summer Activation 2026' all coexist.
search_subfolder now strips every non-alphanumeric character from
both the search input and the folder names before substring match,
so:
"kids drop 1" → matches "1_CFUL263C01C_Kids drop1"
"Spring 2026" → matches "4023 Spring 2026"
"winterfilm" → matches "1_WA20263C01 Winter Film"
Form label/placeholder updated to "Campaign Title" with a hint that
spaces/underscores/hyphens/case are all ignored.
The previous search_subfolder implementation paginated the entire
parent folder before falling back to Box's indexed search API. With
the campaigns folder containing thousands of children, this exceeded
even the new 5-minute background-thread cap and surfaced as 'Search
timed out after 5 minutes' to the user.
Now:
1. Hit the indexed search API first (~1-2s typical, even on huge
parents) — returns immediately on a match.
2. Fall back to a streaming enumeration only for fresh folders Box
hasn't indexed yet (~10 min latency window). Capped at 60s wall
clock so we don't loop forever on a missing campaign.
Also improves the not-found error message to mention the indexing
latency caveat — handles the otherwise-confusing case where a freshly-
created campaign folder isn't searchable for a few minutes.
- /api/search-campaign now kicks off a background thread and returns
immediately. The browser polls /api/progress/<session_id> and fetches
the cached result via the new /api/search-campaign-result/<session_id>
endpoint when complete. Box folder enumeration on a not-found campaign
was taking >30s, exceeding the GCP load balancer's response timeout
and surfacing as 'stream timeout' (not valid JSON) to the user.
- Result cached for 10 min via the existing reporting result_cache
(filesystem-backed → safe across gunicorn workers).
- Form label/placeholder/hint updated: tool accepts a campaign NUMBER,
not a campaign name. Placeholder shows '1993857' instead of
'1011A Spring SS2025'.
Video QC:
* _extract_locale_from_filename now also handles the suffix form
..._XX-yy.ext (case-insensitive both sides), so DOOH/OOH-style
adapt filenames like ..._ES-es.mp4 unblock the price_currency
check instead of skipping with "could not extract locale".
* Batch results page expires the SQLAlchemy session at the top of
the route so the post-completion reload sees committed reports
even when it lands on a different gunicorn worker than the one
that wrote them. Reload delay bumped 1s → 2s for margin.
* visual_quality prompt now passes the filename's market+language
to the LLM and tells it the on-screen copy should be in the
localized language, not the source-language guideline copy.
Stops Spanish-market videos being flagged as "language mismatch
with English campaign guidelines".
Printer Check:
* regions.json rewritten to cover all 10 H&M regions (AME, CEU,
NEU, GCN, IND, SHE, SEU, EEU, EAS, Franchise) with default-all
groups. Two judgement calls vs the screenshot: kept TR for
Turkey (TK is Tokelau in ISO and would break filename matching)
and BR for Brazil (every other code is 2-letter ISO).
Campaign codes:
* New core/utils/campaign_code.py is the single source of truth.
Matches both the legacy 4-digits-plus-optional-letter (1013A,
4116) and the new 11-char alphanumeric with year at positions
5-6 (CFUL263C01D). All four prior parser sites now import from
this helper.
Video Master:
* BOX_CAMPAIGNS_FOLDER_ID switched 156182880490 → 133295752718
(same root the Reporting tool uses). Updated config.py default
and all three .env example files.
* Match page now shows which Box folder the search runs against
(with a clickable link), and on a not-found error explains what
was searched for so missing-campaign cases are self-diagnosable.
The previous in-memory dict only worked with a single gunicorn worker.
With workers=2 in gunicorn_config.py, the async-search worker stored
the result in its own process memory while the dashboard request
landed on the other worker ~50% of the time — cache miss → fell
through to a synchronous Box fetch → exceeded the GCP load
balancer's 30s timeout, returning "stream timeout" to the user even
though the search itself succeeded.
Now stores cache entries as pickled files at storage/cache/<key>.pkl,
shared across workers via the existing volume mount. Atomic writes
via tempfile + os.replace. TTL still 30 minutes. Public API
(cache_set/get/delete/cleanup) is unchanged so call sites in
reporting/routes.py continue to work.
Lifted JWT-cookie auth pattern from the AI QC sibling project:
core/auth/middleware.py validates Azure AD JWTs and stores them in
an httpOnly cookie (hm_aiqc_auth_token). Tenant membership is
enforced by JWTValidator's tid check, which is sufficient for the
tenant-wide access policy chosen for this project.
templates/login.html now drives an MSAL.js popup that POSTs the
ID token to /auth/login. base.html exposes Azure config to all
pages so the logout button can also clear the MSAL session.
app.py's @before_request now checks the JWT cookie and exposes
g.user; modules read user identity via core.auth.current_user_email
so usage logs and created_by columns now record the signed-in
user's email rather than a session value.
Legacy username/password code removed: top-level auth_middleware.py,
jwt_validator.py, deploy/generate_password.py.
Aggregate box_import reports by job_number in SQL instead of fetching
the most recent 100 rows and grouping in Python. The row-level LIMIT
hid older jobs whenever one job's rows filled the window.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Folder discovery groups files by version (V1, V2, ...); only the highest
version per master/adapt is matched. Lower versions are reported as
"superseded" so users can see what was skipped.
- Matching is now an asymmetric 3-pass cascade per adaptation:
Pass 1: masters of same duration (±0.5s) — pHash + AKAZE
Pass 2: masters strictly longer than the adapt — pHash + AKAZE
(shorter masters can't have produced the adapt; never compared)
Pass 3: AI Vision on same-duration / different-resolution masters,
triggered only when Passes 1 and 2 find nothing (covers crops).
- AI Vision default switched from gpt-4o to gemini-2.5-flash (~10x cheaper)
and re-enabled in CampaignMatcher.
- Master temp files now persist for the whole run so Pass 3 can re-read
frames; cleanup still happens via shutil.rmtree at end of run.
- Report shows a "Resolved at" badge per match (Pass 1/2/3) and a new
Superseded Files section.
- New /video-master/report/<id>/download endpoint serves the saved HTML
with attachment headers; Download buttons added to results.html and
view_report.html.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A. Excel upload — /campaigns/pricing/upload now accepts .xlsx/.xls
alongside .pdf. File picker in the campaigns UI matches.
B. Deterministic Excel parser (openpyxl, no LLM) — looks for H&M-style
mastersheets:
- 'MPC Prices' sheet -> flat list of {product_id, language, country,
price, currency, product_name} entries (this is the gold mine).
- Regional sheets (AME/CEU/EEU/...) -> formatted prices per locale
used to derive currency symbol, position, decimal/thousands
separators. Skips OLD/COPY sheets.
Verified against the attached 1013A mastersheet: 448 price entries
across 7 products x 74 locales, 139 locale format entries.
Parser lives in modules/campaigns/pricing_parser.py alongside the
existing PDF path (which now also returns the structured form with
empty _prices).
New lookup shape stored in PricingReference.parsed_data_json:
{"_format": {"en-US": {currency_code, symbol, position, ...}, ...},
"_prices": [{product_id, language, country, price, currency,
product_name}, ...]}
Legacy flat {"<code>": {...}} is still recognised (treated as _format
only) for backwards compatibility with the legacy global JSON import.
Model helpers added:
- PricingReference.get_format_map()
- PricingReference.get_prices()
to_dict() now reports price_count alongside entry_count.
C. Upgraded price_currency_check.py — when a pricing reference with
_prices is attached, the check runs a deterministic comparison:
detected price(s) -> normalize (_normalize_price handles '$49.99',
'39,99 €', 'CHF 49.95', '1.234,56', 'Rs. 2,799', '13 995 Ft', '349,-',
'0.999.000'...) -> compare with tol=0.005 against the expected
per-locale rows. LLM-based campaign-sheet fallback only runs if no
_prices are present (legacy PDF reference or has_pricing campaign
presentation).
D. Video QC price check — new _run_price_check step in the executor.
Parses filename (Market_lang_CampaignNum_... -> 'lang-Market' locale),
detects prices across frames via the same Gemini/GPT-4o path the
other checks use, then deterministic-validates against the attached
pricing reference. Skipped if no pricing ref, unknown locale, GEN/CEN
markets, or no price visible in video.
Overall video score now uses weighted mean of active (non-skipped)
checks (visual_quality w=50, censorship w=50, price_currency w=30)
instead of the hardcoded 50/50 split — so skipping any one check
falls through cleanly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The "Global Pricing Reference" is no longer a single file at
storage/reference/global_pricing.json. Pricing references are now
first-class DB rows (PricingReference model), uploadable as a library
in the Campaigns tab and selectable per-run alongside the campaign
presentation dropdown on the HM QC and Video QC configure pages.
New:
- core/models/pricing_reference.py — PricingReference model: id, name,
pdf_filename, pdf_path, parsed_content, parsed_data_json, status,
created_at/by. get_lookup() deserializes parsed_data_json; to_dict()
powers the dropdown API.
- /campaigns/pricing/upload — creates a PricingReference row, saves PDF
under storage/pricing_references/<id>/, kicks off background parse.
- /campaigns/pricing/<id> DELETE, /campaigns/api/pricing/list,
/campaigns/api/pricing/status/<id>.
- Campaigns index: "Pricing References" table card (mirrors the
presentations card) + upload form with optional name field.
Changed:
- pricing_parser: parse_pricing_pdf_to_dict returns (dict, raw_text);
new parse_pricing_reference(id) runs the parse against a DB row and
sets status to ready/error. Legacy file-based path removed.
- QCExecutor and VideoQCExecutor accept pricing_reference_id; load the
row into context['pricing_reference']={id, name, lookup}.
- BatchQCExecutor and BatchVideoQCExecutor thread pricing_reference_id
through to per-file executors.
- price_currency_check._validate_currency reads context instead of the
disk file; returns 'skipped_no_reference' if no ref attached.
- HM QC + Video QC /execute and /execute/batch routes pass
pricing_reference_id from the JSON payload.
- Configure templates for HM QC and Video QC add a second dropdown
"Pricing Reference (Optional)" loaded from /campaigns/api/pricing/list.
Backwards compatibility:
- app.py: on startup, if storage/reference/global_pricing.json exists
and the pricing_references table is empty, import it as a
"Default (legacy global)" PricingReference row so existing installs
keep a valid reference attached (user can pick it at configure time).
- config.py: retains GLOBAL_PRICING_{PDF,JSON}_PATH for the legacy
importer; adds PRICING_REF_STORAGE_PATH for the new per-row storage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mirrors the existing HM QC batch pattern so Video QC now supports
queueing and processing multiple videos from a single upload.
New:
- batch_executor.py — BatchVideoQCExecutor, sequential processing
(gc.collect() between videos, cooldown between batches), stamps
a shared batch_id into each report's metadata_json.
- /video-qc/execute/batch — kicks off a BatchVideoQCExecutor thread.
- /video-qc/results/batch/<session_id> — batch summary card, per-file
list (filename, score, status, view/download), ZIP download link.
Reuses results.html with is_batch flag.
- /video-qc/report/<id>/download, /video-qc/report/batch/<id>/download
(ZIP), /video-qc/report/batch/<id> DELETE.
Changed:
- VideoQCExecutor accepts batch_id; writes it into metadata when set.
- /video-qc/upload accepts multi-file (request.files.getlist('files'))
with single-file fallback; returns is_batch/filenames/file_count.
- Upload template: drag-and-drop list UI (same pattern as HM QC upload).
- Configure template: shows file count + list, swaps button text and
POST endpoint based on file_count; redirects to results/batch when
batch, results when single.
- Video QC index uses QCReport.get_recent_grouped to render "Batch
Reports" (collapsible per-batch table) + "Individual Reports".
Post-run destinations:
- 1 file -> /video-qc/results/<session_id> (unchanged)
- N files -> /video-qc/results/batch/<session_id> (batch summary +
list of reports from the run)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- video_qc/executor.py: escape braces in JSON example blocks inside
f-string prompts (visual_quality, censorship). Unescaped { } made
Python parse the example as format specifiers, raising
"Invalid format specifier ' 85, ..." and failing execution.
- reporting/routes.py: history_dashboard now passes reports=parsed_reports
(matching the live dashboard route) and attaches friendly_checks per
report. Previously passed parsed_reports=friendly_reports, a kwarg
the template does not consume, leaving the Parsed Data View accordion
empty and breaking the "View Details" scroll-to-file links.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
UsageLog now records input_tokens and output_tokens separately and costs
each side at its real rate. The old single 'blended' rate underpriced
input-heavy workloads (vision/QC) and overpriced output-heavy ones.
COST_PER_MILLION_TOKENS rebuilt against the live OpenAI, Gemini and
Anthropic pricing pages (GPT-5.4 family, GPT-4.x, o4-mini; Gemini 2.5
Pro/Flash/Flash-Lite + 1.5 legacy; Claude 4.7/4.6/4.5 + 3.x legacy).
Unknown models now warn instead of silently defaulting to $5/1M.
Adds idempotent ALTER TABLE migration on startup so existing SQLite DBs
pick up the new columns. Dashboard + API surface the input/output split.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Show job number in batch header instead of just "Batch <date>"
- Add delete batch button (trash icon) that removes all reports + files
- New DELETE /hm-qc/report/batch/<batch_id> route
- Unified batch results view: always renders from DB reports (not
ephemeral progress tracker data), so the view is identical whether
you just completed a batch or navigated back from another tab
- Include thumbnails in batch results per-file rows
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Still OOM after 7 files despite sequential processing. Root cause:
Python's allocator doesn't return freed memory to the OS, so image
buffers accumulate across files until the OOM killer strikes.
Fixes:
- Reduce LLM image max size from 2000px to 1200px (64% less RAM per
image, still sufficient for vision analysis)
- Always close PIL images immediately (not just when opened locally)
- Replace ThreadPoolExecutor with simple sequential loop + gc.collect()
after each file to force memory reclamation
- Switch gunicorn to gthread (2 workers x 2 threads) for better
request concurrency without extra memory overhead
- Add max_requests=200 to auto-recycle workers and release accumulated
memory
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Worker was SIGKILL'd by OOM killer during batch QC (18 files). Fixes:
- Reduce MAX_CONCURRENT_FILES from 2 to 1 (sequential processing)
- Reduce gunicorn workers from 4 to 2 (less memory contention)
- Explicitly close PIL images after thumbnail generation
- Close BytesIO buffers and PIL images after base64 encoding
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Embed asset thumbnail (base64) in HTML report header
- Add view/download buttons to batch results per-file rows
- Add download ZIP and consolidated report buttons to batch results
- Add view/download buttons to upload page recent reports table
- Add download button to individual reports on index page
- New POST /hm-qc/report/consolidated route: merges selected reports
into a single downloadable HTML with summary table + embedded reports
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ThreadPoolExecutor workers don't inherit the parent thread's Flask app
context, causing "Working outside of application context" errors during
batch QC execution. Pass the app instance into BatchQCExecutor and wrap
each child thread's work with app.app_context(). Also ensure the
progress_sessions table is created on fresh databases.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Video QC: Switch to Google Gemini direct video analysis as default (OpenAI frame grid fallback)
- HM QC: Group reports by batch with collapsible sections, ZIP download per batch
- HM QC: Generate asset thumbnails (150px) displayed in report listings
- Speed: Remove artificial delays, add ThreadPoolExecutor(2) for parallel batch processing
- Price detection: Improved prompt with country context, detect all prices, increased text limit
- New Printer Check module: CSV-to-PDF cross-referencing ported from CrossMatch Rust app
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Export/download links in reporting dashboards used hardcoded absolute
paths (e.g. /reporting/export/html/...) which bypassed the reverse
proxy SCRIPT_NAME prefix (/hm-ai-qc-report), causing "No file" errors.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Strip markdown code fences from LLM response before JSON parsing
- Log raw response and parsed result for debugging
- Show warning with provider/model info when detection fails (instead of silent skip)
- Separate "detection failed" (warning, 70) from "no price found" (skipped, 100)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Accept .xlsx/.xls uploads alongside PDFs in campaigns module
- New parse_campaign_excel() in services.py using openpyxl
- Converts all sheets to structured text (headers + rows) for LLM use
- Upload form now accepts both PDF and Excel files
- Added openpyxl to requirements.txt
Workflow: upload campaign presentation (PDF) + media plan (Excel with
has_pricing checked) for the same campaign ID. The price check will
use the Excel data to validate actual prices per country.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Global pricing parser now explicitly extracts format only (symbol,
position, separators) — ignores actual price values in the reference doc
- Executors load ALL ready documents for a campaign (not just the latest),
combining their content — supports guidelines + media plan side by side
- Campaign context now separates pricing_content (from has_pricing docs)
from general parsed_content (all docs combined)
- Price check uses pricing_content specifically for actual price validation
- Report header shows document count (e.g., "1022B - AW25 Display (2 docs) + pricing")
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Filename check:
- Rewritten to flexibly parse multiple H&M naming conventions
(Display, DOOH, OOH, SOME STATIC, Social, POS, DS)
- Extracts country code, language code, dimensions, campaign number
- Scores based on how much metadata was extracted (not rigid pattern)
- Tested against real filenames: BG_bg, ES_es, NO-no formats
Price/currency check (new):
- Detects prices in images via LLM vision API
- Validates currency against global pricing reference (deterministic)
- Falls back to LLM validation for unknown countries
- Optional campaign pricing sheet validation when has_pricing=True
- Added to profile with weight 30
Profile weights rebalanced: filename 30, quality 40, price 30
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add /report/<id>/download route using send_file instead of broken
static file URL (fixes 404 on Download Report button)
- Add campaign label to HTML report header (Campaign: ID - Name)
- Store campaign_id in report metadata_json for traceability
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix background parsing thread: pass app reference explicitly instead of
trying to access current_app inside the thread (was silently failing)
- Add progress bar with animated stages during upload and parsing
- Add data-id/data-status attributes to table rows for auto-polling
- On page load, automatically poll any pending/parsing rows and update
their status badges in-place (fixes stale "Pending" on tab return)
- Immediately inject new row into table after upload so user sees it
without needing to refresh
- Remove broken _parse_pricing_background function
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Introduces a new Campaigns module for uploading campaign presentation PDFs
that QC checks reference to validate assets against campaign-specific
guidelines (typography, layout, copy, pricing format). Also adds a global
pricing reference system that maps country codes to currency symbols and
formats for deterministic price/currency validation.
- New CampaignPresentation model + campaigns blueprint with CRUD routes
- PDF parsing via LlamaParse (text + multimodal page images)
- Global pricing PDF parsed into structured JSON lookup
- Campaign context injected into both image and video QC executors
- Quality checks enhanced with campaign guidelines in LLM prompts
- Price/currency check uses global pricing lookup (saves an LLM call)
- Campaign dropdown added to HM QC and Video QC configure pages
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
AKAZE tier needs the actual video file to extract frames, but our
temp-download-and-delete approach means the file is gone by that point.
Perceptual hash (Tier 1) works fine with saved fingerprint data.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
