plain_persian() strips kasra/fatha/damma/shadda/sukun/hamza-above and the written «هی» ezafe, keeping tanwin. Applied
at output, not in data: PlainPersianJSONRenderer runs one byte-level pass over every JSON response (~5 ms per MB
locally), and to_plain() cleans tutor answers including the stream. Rewriting titles in the DB/seed would have broken
question binding by title path and SeedRecord keys. Tests updated to expect the cleaned text; config/tests_plain_persian.py.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- content.Subject + Unit.subject, /api/home?subject=, tutor grounded in the chosen book
- 5 new books ingested; ~1150 hand-written questions, islands named by hand
- node_split: islands over 7 lessons split into parts of <=6 (lesson ids and progress kept)
- Lesson.figures hand-authored and visually checked for 653 lessons of the new books;
gen_lesson_figures got --subject and survives network errors
- seed files re-exported
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Health: public gets only ok/error; details need X-Internal-Key. Docs/schema: admin session only.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
accounts.0008 drops the columns and refuses to run while guest rows exist.
Docs describe the current state; tests that only proved guest behaviour are gone.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- Admin: refund/cancel/mark-resolved/Bazaar-link actions need change permission; question and
curriculum import need add/delete; content browser needs view_question; guest device ids are
shown to superusers only.
- Accounts: registering a guest's phone revokes the guest's earlier sessions; self-deletion keeps
Redemption, UserOffer and the phone's OTP rows; the public demo guest cannot be deleted
(DEMO_GUEST_DEVICE_IDS); display-name contact filter covers any separator and social handles.
- Payment: callback without Authority is 400; a second Bazaar purchase on a paid order is 409;
orders snapshot credits/hearts/days at creation (payment.0011).
- Assistant: a broken diagram no longer makes a streamed answer free; a late off-topic marker
refuses the answer (not stored, not cached); a disconnect saves the partial answer; only
KNOWN_COLLECTIONS are searched.
- Infra: secret-key file is created atomically (workers no longer split keys); no default
INTERNAL_SERVICE_KEY; Kavenegar key and Bazaar token kept out of error logs; Swagger/ReDoc
pinned; schema throttled; docker compose bound to 127.0.0.1; non-ASCII digit ids are 400;
ticket text capped at 5000.
- docs/security-before-production.md: remaining open items 24-31 and the guard tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- seed_content no longer undoes admin work: a SeedRecord ledger (content/seed_sync.py)
stops deleted/renamed/re-worded content from being recreated and keeps admin-edited
lesson text; new rows in the seed files still arrive (migration content.0010)
- plan list cache is invalidated on Plan save (receiver used a stale 'v1' key)
- token refresh per-IP cap 30/hour -> 600/hour (TOKEN_REFRESH_IP_RATE)
- a failed gateway result only changes a pending order (forged NOK no longer
overwrites refunded/canceled)
- tutor gets the correct answer only after the student has answered
- support reply notification moved to a TicketMessage post_save signal
- global daily cap on paid AI questions (AI_GLOBAL_DAILY_MAX, default 3000)
- Bazaar verify uses only the order's SKU; bazaar orders require a SKU
- /api/health warns on the committed signing key; deploy.sh saves local commits;
400 for bad ?unit= / ?period=; account delete keeps the referrer's Conversion
- earlier today: no stub answers outside DEBUG, ?grade= removed, logged fallbacks
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /api/ai/ask: the stub answer is reachable only under DEBUG or
RAG_FORCE_MOCK. Without RAG keys on the server it used to return 200 with
"this is a stub, set OPENAI_API_KEY" and spend the student's quota; now 503.
- /api/home: the `grade` alias is gone and sending it is a 400 instead of a
silent fallback to the user's own grade. No client sent it.
- activate_subscription logs ERROR when the paid order's plan no longer exists
(it silently became 30 days).
- resolve_grade_title logs instead of `except: pass`; curriculum_io no longer
reads an `intro_text` key no file contains.
- CLAUDE.md: "No silent fallbacks" rules for future agents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ZWNJ normalisation in screen() replaced ZWNJ with itself, so
'نادیدهبگیر' passed, and self-harm sentences typed with a space or
without ZWNJ ('می خوام بمیرم', 'دیگه نمیخوام زنده باشم', 'خود کشی')
reached the model, which answered with the plain math-only refusal
(one live call, 0.018 cents). Self-harm is now matched on the text with
all spaces removed; the imperative must end the word, so 'نادیده بگیریم'
in a math sentence is no longer an injection.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- packs (20 / 60 / 150 questions for 45,000 / 119,000 / 259,000 toman) never expire and
give 7 / 30 / 60 days of unlimited hearts; old subscription plans stay for existing
subscribers but are no longer on sale
- credit ledger (subscription.QuestionCredit): +N per paid pack (once per order), -1 per
tutor question that reaches the API; the free daily questions are spent first; questions
stopped by the server filter, cached answers and failed answers cost nothing
- unlimited hearts from packs live in HeartsPass, so buying a pack never rewrites a
running subscription
- Message.charge records how each question was paid; GET /api/ai/quota and the ask
responses report free-left-today and credits
- HEARTS_ENFORCED now defaults to on: zero hearts stops answering until refill or a pack
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- profile publicId ("User482917", accounts/public_id.py); admin shows and searches it
- plans API reports aiDailyQuota per plan and freeAiDailyQuota; migration 0007 sets the
temporary prices (monthly 99,000 / yearly 990,000 toman, 5 tutor questions a day) unless
an admin already edited the yearly plan
- tutor tone follows the learner's age (18+ gets an adult tone at any grade); notation
still follows the grade's book; answer cache keyed by tone (prompt tutor-v12-age-tone)
- POST /api/user/delete: phone users confirm with a fresh SMS code (same checks as login,
now shared in verify_phone_otp), guests type "حذف"; payment records are kept, demo
accounts refused
- delete_users --skip-paid
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- remove the family app from the root (separate parent accounts, invite codes, parental
settings); progress.0006 drops its tables, permissions and migration history, and
Profile.parent_phone, the two family notification types and ?userId= on
/api/progress/nodes go with it
- GET /api/user/weekly-report: last 7 days next to the 7 before (estimated minutes,
questions, accuracy, islands, tutor questions) for the "parent report" page
- profile birthDate in Jalali (accounts/jalali.py), ages 5-99, birthYear derived from it
- GET /api/payment/invoices; orders now snapshot list price and plan title
- tutor answers only math: server pre-filter, scope rule in the prompt (tutor-v11),
refusal marker replaced with fixed text; self-harm gets a help line instead
- streamed answers now record token usage (it was always 0)
- daily tutor quota: AI_DAILY_QUOTA_FREE (default 3) vs Plan.ai_daily_quota
- delete_users management command for wiping test accounts across every app
- docs/ai-cost-and-guardrails.md: measured cost per question and guardrail results
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A real chat answered "what if 6 were 4?" with a semicircle problem from the
retrieved book excerpt instead of the parabola asked a moment before. A short
follow-up now carries the previous user question next to it. The catalog
teaches line equations as written, regularPolygon, exteriorAngle, graphs only
through function items and tangents through tangent; fraction words (یکپنجم،
نُهدهم) pull the fraction section.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"What if 6 were 4?" has no shape word, so the diagram hint never went next to
it and the tutor would answer without a new figure. When the previous user
question in the conversation had one, a soft hint (redraw only if this
continues that problem) and the same catalog sections now go along.
tools/fake_openai.py stands in for GapGPT locally so the whole tutor path can be
exercised without credit: 12 scripted answers across grades 6-12 went through
streaming, block splitting, repair, storage, history and rendering in the real
chat screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Diagrams: 14 kinds drawn by the app (plane with constructed points, graphs with
computed roots/extrema/tangents, number lines with intervals, solids, charts,
tables, Venn, mapping, trees, fractions, grids, groups, clock). The catalog is
split by topic and only the matching sections go next to a question (200-1073
tokens instead of 1854). parse() repairs slips seen on 96 real book problems:
Persian digits in JSON numbers, 2*pi, // comments, bare items, Venn with
Persian set names. Measured: 45/48 valid on the second round, 0/8 on plain
arithmetic; visual review of 83: 64 right, 3 weak, 16 wrong.
Content: 82 new questions so no island has fewer than four; fill_lesson_texts
writes Unit.blurb (58) and Lesson.example/steps (590) from verified book
problems, never the lesson's own quiz problem. seed_content only fills empty
fields, so admin edits survive a deploy. Tedi unlocks at 500 xp.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The model ends a shape question with <diagram>{json}</diagram>; rag/services/diagram.py
splits it off before mathtext, validates it against a whitelist and stores it on
Message.diagram (ai_tutor.0006). It reaches the app in the SSE done event and in
conversation history. Prompt tutor-v6-diagrams; a hint next to the question (shape
words only) raised gpt-4.1-mini from 3/12 to 12/12 diagrams, 0/8 on arithmetic.
progress.0005 adds tedi (600 xp) to the companion catalog.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two benchmark rounds (14 models on 30 problems, then gpt-4o-mini vs gpt-4.1-mini
on 102 problems under three prompt versions, hand-graded): gpt-4.1-mini scored
95-96/102 against 85-90 for gpt-4o-mini, at 0.12 vs 0.04 cents per answer.
The default now lives in settings so the deploy needs no .env edit.
The context note "in the book / is a decimal point, not a fraction" was written
for the upper-grade books but went to every grade, so gpt-4.1-mini computed
1/7 + 1/8 as 3.5. For primary grades the note now says / can be either, and a
reading rule sits next to the question itself (the same rule at the end of
the system prompt had no effect). Slash problems went from 11 to 14 of 17;
three are still wrong.
generate_answer now reports the provider's exact per-call cost, reasoning
tokens and whether the raw output broke the plain-text rule. The benchmark
fixes primary grades being sent with no book context, caches retrieval so
runs compare on identical context, and retries Qdrant timeouts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tutor (prompt tutor-v3-stepwise-plain): plain text instead of LaTeX/markdown,
with rag/services/mathtext.py converting whatever still slips through; full
step-by-step solutions with a check before the final answer; three tones by
grade; the last 6 messages of the conversation go to the model and follow-ups
are no longer served from another user's cache; book solutions the checker
disputes are kept out of the context; a questionId gives the tutor the
practice question and the student's wrong answer. /api/health reports when
the AI credit runs out. rag_model_benchmark compares chat models on verified
book problems (run on the server; credit ran out mid-run).
Map: only lessons that have questions count toward finishing an island,
lessonIds[0] / nextLessonId is always the next lesson to play, and a finished
island stays finished. The Seyar characters (sigma/delta/pi) are seeded.
Audit fixes, each reproduced before fixing:
- the plan is stored on the order, so paying a monthly order after starting
a yearly checkout no longer grants a year
- opening checkout no longer demotes an active subscription; renewals extend
from the current expiry; expired subscriptions are not reported active
- one discount code per order and only before the gateway; a 100% code
completes the purchase; lost Zarinpal confirmations are re-verified
- the unused /api/payment/webhook "mark paid" endpoint is removed
- XP counts only answers recorded on the server
- 6-digit OTP, each code dies after 5 wrong guesses; POST /api/auth/logout
- a guest who signs up keeps their account and progress (test phones excluded)
- affiliate commission on the first real payment; guests excluded
- parents with several children no longer get a 500
- SQLite runs transactions IMMEDIATE, since it ignores select_for_update
- admin login is rate-limited; display names drop phone numbers and links;
userQuestion is capped at 2000 characters
- server-side heart enforcement exists behind HEARTS_ENFORCED (off until the
frontend handles 403 no_hearts)
Not done on purpose: rotating the server's secret key and moving to Postgres.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The frontend plays node quizzes, but a node only completed once every one of its
8-13 lessons was completed separately, so students were stuck after the first
island of each unit. POST /api/nodes/{id}/quiz/complete now completes the node
(any score; score only sets stars/xp) and returns nextNodeId.
Also closes the other backend gaps found while auditing the frontend:
- /api/ai/ask without lessonId now picks the book from the user's grade instead
of skipping retrieval, so the "ask Pi" tab is grounded in the textbook.
- guest sessions return a refreshToken; they used to go silently 401 after 1h.
- lesson_complete throttle 10/hour -> 120/hour (a lesson here is 1-2 questions).
- progressPercent is per grade, not over all 2,617 lessons of every grade.
- a one-year plan (299,000 toman) is created by migration when no plan exists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ExecuteCampaignRealConcurrencyTests had been written off as a known flake
for weeks — about 45% red on a full-suite run. It was not timing noise.
It was reporting a real defect, and the defect is in production code.
execute_campaign looped with .iterator(), which keeps the read cursor on
the user table open for the whole loop — a loop that writes a
CampaignDelivery per user and, on the sms branch, makes a network call to
Kavenegar in between. On SQLite that open cursor locks the table against
any concurrent writer, and the lock is not released until the loop ends.
So _retry_on_sqlite_lock could never recover from it no matter how long
it waited: widening its window from 0.5s to 2s (also done here, since the
linear backoff was too tight regardless) moved the failure rate not at
all on its own.
Both branches now materialise the ids before the first write. After the
change: 20 consecutive runs of the test itself and 3 full-suite runs, all
green.
This is not only a test fix. While we are on SQLite the same pattern
produces the same lock in production, where it surfaces as "database is
locked" on unrelated requests that happen to land during a campaign run.
Also in this pass
-----------------
- CLAUDE.md carried three sections each claiming to be the current state,
with three different question counts (1,441 / 1,783 / reality 2,463).
There is now one current-state section and the older two are labelled
as history, so whoever picks this up next is not reading stale numbers
as fact. Three other figures that had drifted are corrected against the
database: the verdict split, the figure-gap share (881/~42% is now
508/~14% after the five books were re-read), and the node coverage.
- Dropped four genuinely unused imports. The remaining pyflakes hits are
the signal registrations in apps.py, which are imported for effect.
- .env.example suggested gpt-4o for vision; gpt-4.1-mini measured both
cheaper and more accurate and is what the server actually runs.
Audited all 2,463 questions across every grade: no question with the
wrong number of options, none with zero or two correct answers, no empty
or duplicated option, no orphan lesson or node, no duplicate question
inside a lesson.
458 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Content
-------
Every node in grades 8, 9, 10, 11 and 12 now holds a full four-question
quiz — 121 more questions, filling the two حسابان دوازدهم nodes in
فصل ۵: کاربردهای مشتق that were empty and the two ریاضی هشتم nodes that
were short. The repo's own seed guard was already failing on those two
empty nodes; it passes now.
Grades 3, 6 and 7 are the other team's books and were left alone. 39 of
their 145 nodes still hold fewer than four questions, but none is empty,
so every node on the map still plays.
An audit over all 1,593 questions in grades 8–12 found no structural
fault: no question with the wrong number of options, no question with
zero or two correct answers, no empty or duplicated option, no leftover
LaTeX, and no two identical questions inside one lesson.
Contract
--------
Three endpoints were missing or malformed in the OpenAPI schema, which
matters more than it looks: that schema is the contract the frontend
generates its client from, so a view missing there is a view the other
repo cannot call.
- NotificationMarkReadView was dropped from the schema entirely, because
spectacular could not guess a request body for its PATCH and gives up
on the whole view rather than the one field. It declares request=None.
- PaymentRedirectView was dropped for the same reason; it answers with a
302 and no JSON body, which it now says.
- /api/health and /api/health/ collided on operationId and were emitted
as health_retrieve and health_retrieve_2. The slash variant exists only
for a monitor that calls it that way, so it is excluded from the schema.
`spectacular` now runs clean — no errors, no warnings.
Tidying
-------
Condensed the two longest comment blocks (18 and 14 lines) in
rag_verify_solutions and settings, keeping the measurement and the rule
and dropping the retelling. README gains the bulk question-authoring
workflow; the playbook's subagent section now points at the real tool
instead of describing the steps by hand.
458 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/extraction-playbook.md: چهار تصمیمی که تفاوتِ چند برابری در هزینه و کیفیت
ساختند، با عددهای اندازهگیریشدهی واقعی نه تخمین.
سرفصلها: انتخابِ مدل از روی اندازهگیری (gpt-4.1-mini هم دقیقتر از gpt-4o بود
هم ۳٫۵ برابر ارزانتر، و gemini برخلافِ جدولِ قیمت سه برابر گرانتر)، قاعدهی
self-contained، تعمیرِ بکاسلشِ LaTeX، محافظِ شکل، و ساختِ سوال با subagent
بهجای API.
وجهِ مشترکِ هر چهار مورد: هیچکدام خطا تولید نمیکردند — نه در لاگ، نه در تست.
فقط دادهٔ کمتر یا بدتر میساختند. برای همین توصیهی محوریِ فایل این است که قبل
از اعتماد به هر ران انبوه، نمونهای ۴۰-۵۰تایی با چشم خوانده شود.
تستهای rag هم اینجا میآیند: بازخوانی، تعمیرِ JSON، و محافظِ شکل.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
محافظِ «وابسته به شکل» یک regexِ تخت بود (شکل|جدول|روبهرو|مقابل|زیر را) و
مسئلهها را بدونِ هیچ فراخوانی کنار میگذاشت. فاجعهبار گشاد بود:
«زاویههای روبهرو در لوزی برابرند؟» ← «روبهرو» یعنی مقابل، نه شکل
«کسرهای زیر را گویا کنید: ۲/√۳» ← داده همانجاست
«ظرفی به شکل مخروطی با شعاع ۴» ← «به شکلِ» یعنی شبیهِ
اندازهگیری روی نمونهی ۵۰تایی از کتابهای بازخوانیشده: ۸۴٪ از ردشدهها واقعاً
قابلِ قضاوت بودند — یعنی ~۷۸۴ مسئلهی سالم بیصدا از چرخهی ساختِ سوال بیرون
مانده بود، بدونِ هیچ خطایی در هیچ لاگی.
قاعدهی تازه دو شرط دارد و هر دو باید برقرار باشند: ارجاعِ صریح به چیزی بیرونِ
متن، و نیامدنِ هیچ توصیفی بعد از آن. شرطِ دوم همان چیزی را میگیرد که قاعدهی
self-contained تولید میکند: «با توجه به شکل مقابل که شامل {a,1} است».
اثرش روی دادهٔ واقعی: unjudgeable از ۱٬۴۴۱ به ۵۰۸ و agrees از ۷۹۱ به ۲٬۴۶۶.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ingestِ دوباره جایگزین نمیکرد: qdrant_point_id یک uuid4 تازه به ازای هر ردیف
است، پس خواندنِ دوبارهی یک صفحه نسخهی تکراری کنارِ نسخهی قدیمی میگذاشت و
جستجو نمیتوانست تشخیص بدهد کدام کدام است.
--refresh {figures,unjudgeable,both,all} و --refresh-pages اضافه شدند.
سه تصمیمِ طراحی که عمدیاند:
۱. پاکسازی صفحهبهصفحه است و درست پیش از خواندنِ همان صفحه، نه یکجا از اول.
وسطِ تست یک crash همین را ثابت کرد: فقط یک صفحه خالی ماند (۲۷۹→۲۷۶) و بقیهی
کتاب دستنخورده. پاکسازیِ یکجا همان crash را به ۸۱ صفحهی خالی تبدیل میکرد.
۲. اول Qdrant، بعد دیتابیس. شناسهی نقطه فقط در ردیفِ DB است، پس ترتیبِ برعکس
یعنی شکستِ Qdrant نقاط را برای همیشه بیصاحب میگذارد — نه قابلِ یافتن، نه
قابلِ پاک کردن.
۳. order از Max گرفته میشود نه count: بازخوانی ردیفها را از وسط پاک میکند و
count عددی میداد که با orderهای موجود تصادم داشت.
تخمینِ هزینه حالا از اندازهگیریِ واقعیِ پنلِ صورتحساب میآید نه از جدولِ قیمت،
و وعدهی «با Batch API نصف» حذف شد چون GapGPT اصلاً /v1/batches ندارد.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
مدلِ vision در متنِ ریاضی LaTeX مینویسد (\frac، \sqrt، \vec) و بکاسلشِ تنها
در JSON نامعتبر است، پس json.loads کلِ صفحه را دور میریخت — دقیقاً صفحههای
پرفرمول، یعنی باارزشترینها. روی حسابان یازدهم/دوازدهم چهار صفحه هرچه تلاش
شد «خطادار» ماندند تا همینجا پیدا شد.
⚠️ نسخهی اولِ تعمیر خودش تله داشت: \f، \t، \b، \r و \n از نظرِ JSON معتبرند،
پس «سالم» رد میشدند و \frac به formfeed+«rac» و \theta به tab+«heta» تبدیل
میشد — خرابیِ بیصدا بهجای خطای بلند. قاعدهی تفکیک: دستورِ LaTeX همیشه با
حرفِ ASCII ادامه مییابد، ولی escapeِ واقعی با رقم، فارسی یا پایانِ رشته.
کنارش: پایههای یازدهم و دوازدهم به persona و تشخیصِ پایه اضافه شدند («دوازدهم»
پیش از «دهم» بررسی میشود چون زیررشتهاش را دارد)، سقفِ max_tokens از ۳۰۰۰ به
۶۰۰۰ رفت چون قاعدهی self-contained خروجی را بلندتر کرده، و حسابان یازدهم و
دوازدهم با نگاشتِ فصلبهصفحه ثبت شدند.
«حسابان» به MATH_SUBJECTS اضافه شد: بدونِ آن retrieval این کتابها را «درسِ
متنی» میدید و بهجای chapter روی lesson_number فیلتر میکرد — بیهیچ خطایی
همیشه صفر نتیجه.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
سه کارِ بههمپیوسته که از یک کشف شروع شدند: راهحلهایی که vision از کتاب
درآورده گاهی غلطاند، و این خطا در بازیابی دیده نمیشود — فقط وقتی کسی به
`solution` اعتماد کند خودش را نشان میدهد.
۱) بازبینیِ راهحلها (rag_verify_solutions)
هر chunk دوباره حل میشود و با کتاب مقایسه. نتیجه سهحالته است، نه بله/خیر:
اجرای اول با بولی ۱۹٪ خطا گزارش کرد؛ خواندنِ دستیِ نمونهها نشان داد ۳۵٪ از
آنها مسئلههایی بودند که به شکلِ کتاب ارجاع میدادند و اصلاً قابلِ قضاوت
نبودند، و ۲۴٪ موردی که «۳/۱۰۴» را کسرِ ۳÷۱۰۴ خوانده بود نه عددِ ۳٫۱۰۴. با دو
حالت، «نمیدانم» ناچار به «غلط است» ترجمه میشد.
با سه حالت و پرامپتی که این دو تله را میشناسد: ۲٬۱۱۹ chunk بررسی شد،
۱٬۱۱۲ درست، ۱۲۳ غلط (۵٫۸٪)، ۸۸۱ غیرقابلِ قضاوت. آن ۸۸۱ رایگان برچسب خوردند
چون پیشفیلتر قبل از هر فراخوانی کنارشان گذاشت.
`solution` هرگز بازنویسی نمیشود — کتاب باید کتاب بماند. برچسب در payloadِ
Qdrant هم مینشیند، ولی با نامی که `retrieval._payload_text` نمیخواند: یک
نویسه هم به متنی که به مدل میرسد اضافه نشد.
نمونهی خطاهای واقعیِ کتاب که پیدا شد:
BMI = ۱۰۰/(۲٫۷۴)² کتاب: ۳۳٫۲ درست: ۱۳٫۳
۱۶ آبی + ۲۴ صورتی کتاب: «کل ۵۰» درست: ۴۰
بزرگترین عددِ ۴رقمیِ بیتکرار کتاب: ۸۵۳۲ درست: ۹۸۷۶
g(x) = x³ − ۲ کتاب: «نمایی» درست: چندجملهای
۲) سوالها فقط از راهحلهای تأییدشده
۱٬۱۹۹ سوال در content/seed/questions.json، همه از chunkهایی که برچسبِ
«درست» گرفتهاند. هر شش پایه برنامهی درسی دارند (۱٬۹۲۸ درس) — هشتم و دهم
تازه ساخته شدند.
۳) GET /api/nodes/{id}/quiz
واحدِ بازی در اپ «مبحث» است ولی سوالها به «درس» بستهاند، و درس اینجا
ریزترین دانه است: یک نوعِ مسئله، نه یک جلسه. نتیجهاش این بود که اپ برای یک
آزمونِ چهارسوالی باید دهها درخواست میزد. این اندپوینت از همهی درسهای مبحث
نمونه برمیدارد، حداکثر یکی از هر درس تا آزمون روی یک مهارت گیر نکند.
درهمریزی با بذرِ nodeId:userId پایدار است: رفرشِ صفحه همان آزمون را میدهد،
نه مجموعهی تازهای که پیشرفت را بیمعنا کند. `questionCount` جدا از
`requested` میآید تا مبحثِ کمسوال وانمود نکند کامل است.
──────────
اصلاحاتِ جانبی که سرِ راه لازم شدند:
• QDRANT_TIMEOUT_SECONDS: کلاینت بدونِ timeout ساخته میشد و پیشفرضِ ۵
ثانیهایِ کتابخانه ingestِ کتابِ هشتم را روی اولین فراخوانی کشت.
• پاکسازیِ نقاطِ نیمهنوشته: upsert با wait=True ممکن است نوشته باشد و فقط
جوابش تایماوت شده باشد؛ مسیرِ جبران فقط ردیفهای DB را پس میگرفت و هر
retry یک نسخهی تکراری در Qdrant جا میگذاشت (۸ مورد در پایه نهم، پاک شد).
• تشخیصِ سقفِ حساب: «insufficient user quota» با فاصله میآید نه زیرخط، و
محافظِ اول نمیگرفتش — یک اجرا ۷۷۰ فراخوانیِ محکومبهشکست زد.
• پرامپتِ vision حالا میخواهد محتوای شکل داخلِ متنِ مسئله نوشته شود. ۴۲٪ از
محتوای موجود به شکلی ارجاع میدهد که در متن نیست؛ مدل آن را دیده بود ولی
از او خواسته نشده بود بنویسدش. کتابهای موجود قبل از این قاعده وارد شدهاند.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
بررسیِ collectionها بعد از ingestِ پایه نهم نشان داد Qdrant ۲۸۷ نقطه دارد ولی
دیتابیس ۲۷۹ ردیف. هشت نقطهی اضافه، همه از صفحهی ۱۲۵ — همان صفحهای که دو بار
تایماوت داد و بار سوم گرفت. سه مسئله سه نسخه داشتند و دو نقطه هم بیصاحب بود.
علت: `upsert` با `wait=True` ممکن است نوشته باشد و فقط جوابش تایماوت شده
باشد. مسیرِ جبران فقط ردیفهای DB را پس میگرفت، نه نقطهها؛ و چون هر اجرای
دوباره `qdrant_point_id` تازه میسازد، هر retry یک نسخهی تکراری اضافه میکرد.
اثرش در بازیابی دیده میشود نه در لاگ: top-k با سه نسخه از یک مسئله پر میشود
و بقیهی متنِ مرتبط بیرون میماند — یعنی دستیار بافتِ کمتری میبیند و هیچکس
هم متوجه نمیشود.
حالا در صورتِ شکست، نقطهها هم پاک میشوند. اگر خودِ پاکسازی هم شکست بخورد
لاگ میشود ولی خطای اصلی پوشانده نمیشود، چون صفحه باید ناتمام بماند تا اجرای
بعدی دوباره سراغش برود.
دادهی موجود هم تمیز شد: ۸ نقطهی بیصاحب پاک شدند و حالا ۲۷۹ = ۲۷۹.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
خطاهای این پروژه خاموشاند: نامِ اشتباهِ collection یا فیلترِ نخورده نتیجهی
خالی میدهد نه استثنا، و دستیار فقط بدتر جواب میدهد. این تستها نگهبانِ
همانجاهایند.
TextbookLayoutTests بررسی میکند رشتهی هر فصل بینِ ingest و retrieval یکی
باشد و بازهی صفحهها پیوسته بماند. IngestBookResumeTests مطمئن میشود
صفحهی بیمسئله دوبار پول نمیگیرد، صفحهی خطادار در اجرای بعد دوباره
سراغش میروند، و upsertِ شکستخورده rollback میشود.
LessonContextResolutionTests نگهبانِ نرمالسازیِ «پایه نهم» → «ریاضی نهم»
است و اینکه مسیرِ gradeId و مسیرِ lessonId به یک collection برسند.
GradeAwarePromptTests جلوی برگشتنِ لحنِ دبستانی برای پایههای بالا را
میگیرد. FollowUpStaysGroundedTests پیامِ دومِ یک گفتگو را هم مستند نگه
میدارد.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
مهمترین خواستهی سینیور این بود که بسنجیم پاسخِ AI واقعاً یاد میدهد یا نه.
برای سنجیدن، اول باید پاسخها قابلِ ردیابی میشدند.
پرامپتها پایهآگاه شدند: تا پیش از این «معلمِ دبستانی» برای پایه نهم هاردکد
بود، هم در تولیدِ پاسخ و هم در vision — یعنی دانشآموزِ نهم لحنِ ابتدایی
میگرفت. PROMPT_VERSION حالا روی هر پیام ذخیره میشود.
هر پاسخ حاشیهنویسی میشود: مدل، نسخهی پرامپت، پایه، collection، مقدارِ
فیلتر، تعداد chunk، توکن، تأخیر و from_cache. فیدبک reason اختیاری گرفت تا
لایک/دیسلایک بعداً قابلِ تحلیل باشد («بشه چک کرد کدوم بهتر بوده»).
ai_feedback_report --by prompt_version نسبتِ رضایت را کنارِ مصرف/هزینه
میگذارد. rag_eval سوالهای واقعی را از مسیرِ واقعی میپرسد، با
--retrieval-only برای اجرای ارزان.
rag/services/lesson_context.py از lessonId تنها پایه و فصل را درمیآورد؛
normalize_grade_title هم اختلافِ «پایه نهم» در دیتابیس با «ریاضی نهم» در
نامِ collection را میبندد — بدونِ آن مسیرِ gradeId بیسروصدا خالی برمیگشت.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
نقشهی فیزیکیِ کتابها (افستِ صفحه، بازهی هر فصل، درسهای داخلِ فصل از
فهرستِ خودِ کتاب) در rag/services/textbooks.py ثبت شد تا عنوانِ فصل از
کتاب بیاید نه از حدسِ مدل — رشتهی فصل باید دقیقاً بینِ ingest و retrieval
یکی باشد وگرنه نتیجه بیسروصدا خالی برمیگردد.
rag_ingest_book کتاب را فصلبهفصل ingest میکند و قابلِ ازسرگیری است:
rag.IngestedPage نشانهی «این صفحه پردازش شد» را نگه میدارد حتی وقتی
صفحه chunkی نساخته، وگرنه هر بار اجرای دوباره بابتِ صفحههای بیمسئله
دوباره پول میگرفت. --plan هزینه را رایگان برآورد میکند.
rag_import_collection نقاطِ موجودِ Qdrant را در LessonChunk میریزد —
فقط میخواند، رایگان و idempotent، برای پایههایی که تیمِ دیگر ingest
کرده بود.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two bugs made self-ingested content useless, both silently:
- _payload wrote `chapter_number` (int) while math retrieval filters on
`chapter` (the full "فصل ۲: کسر" string), so anything we ingest could
never match a search — no error, just permanently invisible content.
- ingest_pdf flattened the vision extraction into one text blob and threw
away type/keywords/generalSolution/problem/solution. LessonChunk is the
source of truth, so rebuilding meant paying for vision again — and the
problem/solution pairs are the input for building lesson questions.
LessonChunk now stores those fields and _payload emits them under the same
keys the other team's collections use, so ours and theirs stay one shape.
Also fixes a regression from the whitenoise commit: the Manifest storage
variant raises when staticfiles.json is absent, which broke all 23 admin
rendering tests (Django forces DEBUG=False in tests, so USE_FINDERS is off)
and would have taken the admin panel down entirely the day DEBUG=0 ships
without a collectstatic run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
rag/services/tutor.py
- deleted: answer_question() had no callers (AskAIView calls
retrieve_context + generate_answer directly) and carried an unused
import.
rag/services/config.py
- docstring: drop the claim that the lesson-filter key is settings-driven
(it is auto-selected per subject in known_collections.py).
.env.example
- drop the leftover note about the removed RAG_QDRANT_LESSON_FILTER_KEY.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
progress/views.py
- GuestAwareReadThrottle (same guest-by-IP pattern as the other apps);
throttle_scope='read' on UserSummaryView, StreakCalendarView,
LeaderboardView.
notifications/views.py
- GuestAwareReadThrottle + throttle_scope='read' on NotificationListView.
- list capped at NOTIFICATIONS_MAX=100 newest (Meta.ordering already
newest-first); unreadCount still counted over the full set.
notifications/tests.py
- 150 rows -> 100 returned newest-first, unreadCount = 75.
rag/views.py
- ChunkListView: explicit order_by('-id') before the existing [:200] cap.
ai_tutor/views.py
- ConversationMessagesView: return the newest 200 messages in
chronological order instead of the whole conversation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Self-contained app; wired into settings/urls in the next commit. Data
stays in the parallel team's live Qdrant — nothing is copied into our DB.
apps.py
- RagConfig.
models.py
- LessonChunk: DB-of-record for one summarized lesson chunk
(original_text + summary + embedding_vector) mirrored as a single
Qdrant point; loose lesson_id/grade_id ints, no cross-app FK.
LessonChunkImage for future multimodal lessons.
migrations/0001_initial.py
- initial tables for the two models.
services/clients.py
- lazy, lru_cache'd OpenAI + Qdrant clients (Rag_backend builds them at
import time and crashes without keys). RagNotConfigured raised when a
key/lib is missing; Qdrant client uses check_compatibility=False for
the 1.19-client / 1.17-server minor version gap.
services/config.py
- rag_is_configured(), collection-name template, grade-title resolution
(RAG_GRADE_TITLE_OVERRIDES -> content.Grade.title), retrieval top-k.
services/known_collections.py
- exact live Qdrant catalog: 6 collection names, math chapter-number ->
exact "فصل N: ..." string tables per (subject, grade), and the
علوم/اجتماعی lesson-number -> title maps. filter_key_for_subject()
picks `chapter` for math, `lesson_number` otherwise;
resolve_math_chapter() maps a plain chapter number (Persian or Latin
digits) to the stored string.
services/embedding.py
- embed_text() pinned to OPENAI_EMBEDDING_MODEL (text-embedding-3-small,
1536d) so vectors match the other team's existing collections.
services/chunking.py
- split_text_into_chunks(), summarize_chunk() (gpt-4o-mini, JSON out);
plain exceptions instead of ninja HttpError.
services/qdrant_store.py
- ensure_collection(), upsert_points(), search(): the lesson filter
matches the value both as str and native, since Qdrant stores
lesson_number as int and chapter as string.
services/retrieval.py
- resolve_target() -> (collection, filter_key, filter_value), reused by
the cache key and debug block.
- retrieve_context() -> (chunks, meta): embed -> Qdrant search scoped to
one lesson/chapter -> top-k. Math payloads (no `summary`) fall back to
generalSolution/problem/solution; a 404 collection returns empty.
services/generation.py
- generate_answer() and stream_answer() (token iterator for SSE);
context-grounded Persian prompt that explains from retrieved text
rather than refusing.
services/pdf.py
- textbook PDF -> page images -> gpt-4o vision problem-type extraction
(adapted from Rag_backend mathapp); returns dicts, no on-disk JSON.
pymupdf/pdf2image optional — missing libs raise RagNotConfigured.
services/ingest.py
- ingest_text() / ingest_pdf(): chunk -> summarize -> embed -> persist
LessonChunk + upsert Qdrant, in one transaction.
services/answer_cache.py
- short-TTL cache of the final answer keyed by
sha1(collection | filter_value | normalized-question).
services/tutor.py
- thin answer_question() (retrieve -> generate) for direct callers.
serializers.py, urls.py, views.py
- POST /api/rag/ingest/text, POST /api/rag/ingest/pdf (HasInternalServiceKey,
throttle scope rag_ingest), GET /api/rag/chunks review list.
RagProviderUnavailable -> clean 503.
admin.py
- LessonChunk admin with an inline for its images.
management/commands/rag_catalog.py
- read-only live Qdrant listing (--values, --json); never writes.
tests.py
- mocked-client tests: ask-with-retrieval, ingest, SSE stream, answer
cache, and rag_is_configured() gating.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
