Handle cashback commission imports, include commissions in analytics with separate investment metrics, and expose commission/version details in the UI.
Made-with: Cursor
SSE streaming added unnecessary complexity and latency due to
buffering issues across Node.js event loop, Nginx proxy, and
Docker layers. Reverted to a simple synchronous request/response
for PDF conversion. Kept extractLlmErrorMessage for user-friendly
LLM errors, lazy-loaded pdf-parse, and extended Nginx timeout.
The for-await loop over OpenAI stream chunks runs synchronously when
data is buffered, causing res.write() calls to queue without flushing.
Add setImmediate yield after each progress event so the event loop
reaches its I/O phase and pushes data to the network immediately.
Add gzip off to Nginx import location — the global gzip on was
buffering text/event-stream responses. Pad each SSE event to 4 KB
with comment lines to push past any remaining proxy buffer threshold.
Add 2 KB padding comment after headers to push past proxy buffer
threshold, enable TCP_NODELAY on the socket, and remove erroneous
chunked_transfer_encoding off from Nginx that caused full response
buffering.
- Move import modal visibility into ImportContext so the Layout
progress pill can reopen the result dialog after the modal was closed.
- Raise LLM progress cap from 85% to 98% and drop the intermediate
-import 88%- SSE event to eliminate the visual stall after LLM finishes.
- Detect LLM context-length errors (n_keep >= n_ctx) and surface a
clear message instead of generic -Временная ошибка конвертации-.
Hardcoded EXPECTED_CHARS (15k) caused progress to stall at ~20-25% for
short statements. Now expected size is derived from input text length.
Also emit an explicit 85% event when the LLM stream finishes, and
throttle SSE events to 300ms to reduce browser overhead.
Replace the synchronous PDF import with Server-Sent Events streaming
between the backend (LLM) and the browser. The user can now close the
import modal and continue working while the conversion runs — a fixed
progress bar in the Layout shows real-time stage and percentage.
pdf-parse@2.4.5 depends on @napi-rs/canvas (native Skia binary compiled
with AVX instructions) which crashes with SIGILL on the Synology NAS CPU.
Version 1.1.1 is pure JavaScript and works on any architecture.
pdf-parse@2.4.5 pulls in @napi-rs/canvas (native Skia binary) which
crashes on import in Alpine containers. Moved to lazy require() so the
app starts normally and pdf-parse loads only when PDF conversion is
actually requested.
- Lazy-load pdf-parse in pdfToStatement to avoid startup crash
- Add libc6-compat, fontconfig, freetype to Alpine runner stage
- Increase npm fetch timeouts in both Dockerfiles for slow networks
- Add connectionTimeoutMillis to pg Pool for faster failure detection
- Добавлена переменная LLM_MODEL в конфиг
- Увеличен max_tokens до 32768 для крупных выписок
- Включен response_format: json_object
- Добавлен скрипт test:llm для проверки подключения к LLM-серверу
- Accept only PDF and JSON files in import modal and API
- Convert PDF statements to JSON 1.0 via LLM (OpenAI-compatible)
- Use multipart/form-data for file upload (multer, 15 MB limit)
- Add LLM_API_KEY and LLM_API_BASE_URL for configurable LLM endpoint
- Update ImportModal to validate type and send FormData
- Add postFormData to API client for file upload
