Repository navigation
ParsRAG v1.1.0
Added
-
Settings in the app can change the options that used to need
.envand a
restart (27 in all), in four groups:- model: context sizes, thinking, retries, timeouts
- retrieval: threshold, depths, keyword search
- documents: upload limits, OCR, image previews
- indexing speed: OCR workers, CPU threads, batches
Each shows its
.envvalue and can be reset to it. Overrides survive
restarts. Ports, allowed addresses, Qdrant, memory limits, and the embedding
model stay in.env. -
Answer sources show the image an excerpt was read from, next to the excerpt.
-
Installed Ollama models are offered in a menu, with their sizes.
-
.env.examplenow has only the model choice active. Every other setting
shows its default, grouped by whether Settings can also change it. -
The README starts with one complete setup using every capability; other ways
to run follow it. -
The app can use the Ollama installed on the computer (Windows, macOS,
Linux) with its existing models and GPU. Set
OLLAMA_DOCKER_BASE_URL=http://host.docker.internal:11434in.envand
start without thelocal-modelprofile. No override file is needed, and the
app container needs no GPU. The README's "Complete run" section covers
each system, GPU memory, and troubleshooting. -
OLLAMA_ALLOWED_HOSTSadds hosts that Ollama URLs may use.
OLLAMA_INTERNAL_HOSTis still read. -
Settings explains Ollama failures ("address not allowed", "did not answer")
in Persian and English. When an API is active, it offers the server's own
Ollama URL. -
Images inside Word and PowerPoint files are cited by number: text read
from an image is its own excerpt with the image's caption, shown as
"paragraph 120 · image 7" or "slide 25 · image 9", and answers say when a
fact comes from an image. Each image excerpt also carries the text that
introduces it (the end of the paragraph before it, or the slide's text). -
Word tables are cited by number and the paragraph they follow
("paragraph 120 · table 3") instead of an internal element index. -
When part of a table is retrieved, the rest of the table comes with it, so
questions about the largest or smallest value see every row.
Improved
-
Questions that ask several things ("… with which key, and which fields …")
are also searched part by part. Each part keeps its best excerpts, and the
model is asked to answer every part. In the acceptance set, 5 of the 6
answers that missed a part became complete, with no regressions among the
questions that were already right. -
Indexing is faster with the same results:
- Pictures in Word and PowerPoint are read several at a time
(OCR_WORKERS). The 96-image final report parsed in 33 s instead of
121 s, with identical text. - Each document is embedded in length-sorted batches.
Indexing that whole report went from about 4.5 minutes to about 3.
- Pictures in Word and PowerPoint are read several at a time
-
OCR reads screenshots better: a second, sparse-text pass adds text that
page layout analysis skipped (status panels, dark UI themes). -
Up to 200 images per Word or PowerPoint file are read (was 30).
-
With several documents, specific questions search all files together, so
the file that holds the answer can provide most excerpts; each other file
still contributes its best match, and passages repeated across files are
kept once. -
Commands and code in one-cell boxes in Word stay with the sentence that
introduces them, so questions about them find them. -
OCR reads text on tinted panels (for example a light-blue status box).
Fixed
[extract_itex]...[/extract_itex]markup that Gemini models leak into answers
is shown as math, and table rows inside a source excerpt render as a table.- Thinking models on Ollama (such as
qwen3.5) no longer return empty
answers. They could spend the whole answer budget on reasoning, so Ollama
is now asked not to think unlessOLLAMA_THINK=1. On a 6 GB GPU with
qwen3.5:4b, an empty reply after 59 s became a cited answer in 7 s. - Reasoning that some APIs put inside answers (
<thought>...</thought>from
Google AI Studio's Gemma 4,<think>...</think>elsewhere) is removed from
answers, streams, conversation titles, and suggested questions. - Transient API failures (HTTP 5xx and 429) are retried up to
MODEL_API_MAX_RETRIEStimes (default 4) with growing delays; Google AI
Studio's free tier fails about one Gemma request in three. - Google AI Studio model ids get the
models/prefix its endpoint requires. - English questions get English answers in Hybrid and open-chat modes; they
were often answered in Persian. Every prompt now names the answer language,
detected from the question. - The service retries startup when Qdrant is briefly unavailable instead of
staying not ready until restarted. - A question is no longer limited to one file just because it shares an
ordinary word with the file name (such as «اول» or «تست»). - An API provider that is out of credit (HTTP 402) is reported as such, with
a hint to top up the account or choose another model.
Security
- Questions with long runs of spaces or many unclosed
@{mentions can no
longer make question splitting or answer-language detection slow (CodeQL
regular-expression alerts).
Changed
- Every backend test runs without skips and without Docker:
PARSRAG_E2E=1 pytestruns the end-to-end tests against an in-memory
Qdrant unlessQDRANT_HOSTnames a server. - Dependencies: FastAPI 0.143, Uvicorn 0.54, python-dotenv 1.2.4, Ruff 0.17,
mypy 2.4, KaTeX 0.19, Vite 8.3.4, Playwright 1.64, lucide-react 1.56,
msgpack 1.2.3 or later.
Upgrade note
- Documents indexed by 1.0.0 keep working. Upload Word and PowerPoint files
again to gain image and table citations, image previews next to sources,
images past the 30th, and the improved OCR of screenshots. - Settings changed in the app are saved in the
app_configvolume and take
precedence over.env; each can be reset to its.envvalue.
Acceptance test
52 detailed Persian questions over 10 real documents (one PowerPoint and nine Word reports, 244 images), each asked in both modes with gemini-flash-lite-latest:
| 1.0.0 | 1.1.0 | |
|---|---|---|
| Documents only: correct answers | 33/52 | 44/52 |
| Hybrid: correct answers | 30/52 | 42/52 |
| Image questions correct (documents only / hybrid) | 5 / 4 of 19 | 12 / 11 of 19 |
| Table questions correct (documents only / hybrid) | 10 / 9 of 11 | 11 / 11 of 11 |
| Answer cites the right document (documents only / hybrid) | 45 / 48 | 51 / 52 |
The 1.0.0 column used google/gemma-4-26b-a4b-it, so part of the difference comes from the model. Questions that ask several things were measured afterwards: 5 of the 6 incomplete answers became complete, with no regressions.
Full changelog: v1.0.0...v1.1.0