Realist
An open-source product listing audit tool for detecting contradictions and inconsistencies among titles, descriptions, specifications, and images in e-commerce listings, using a hybrid architecture that combines deterministic comparison in Python with observations from vision models.
Timeline (last 90 days)
Realist is an open-source product listing auditing tool used to perform quality checks on sellers' existing product listings and discover contradictions or inconsistencies among titles, descriptions, specifications, and photos.
Realist's findings fall into four categories: text_conflict (fields contradict each other), image_mismatch (photos do not match the text), possible_omission (visible content is not stated), and unverified_claim (claims unsupported by materials, not accusations).
Realist splits the work into two parts: pure Python code handles precise comparisons (such as Bluetooth version, weight, color, etc.), and the vision model only returns structured observation results without making judgments.
Because the Python comparison is deterministic and involves no sampling or temperature parameters, the confidence of text conflicts (text_conflict) is fixed at 1.0.
The confidence cap for photo-based color mismatches is set at 0.8, because lighting and image processing can change colors; when the model number digits recognized by OCR are close to but different from the listing, the confidence is 0.5, because OCR can misread characters.
Realist runs the Qwen2.5-VL-7B vision model through a local Ollama server (default qwen2.5vl:7b), and photos are scaled down and stripped of metadata before being sent only to 127.0.0.1, not flowing externally.
Realist's visual audit relies on a three-line VisionBackend abstraction interface, making vision models such as Qwen2.5-VL replaceable components.
Realist uses PaddleOCR as an optional OCR layer to extract text such as model numbers and labels from photos and compare them against the listing; both the model and PaddleOCR use the Apache-2.0 license.
Realist has tested the deterministic engine, including 10 unit tests covering matching listings producing zero findings, constructed conflicts being caught, unit equivalences not being flagged, specification-supported claims not being flagged, specification-refuted claims being caught, and HTML export escaping malicious input, among others.
Realist has not yet tested the real observation accuracy of photo inspection, because it uses handwritten observation results; it also has not yet built an annotated evaluation set with precision, recall, and false positive rates, nor a pure-model baseline.
4 more timeline events
All Facts (14)
The Realist project is hosted in the GitHub repository rafayet-git/realist and can be started with streamlit run app.py; photo analysis is optional and requires installing Ollama and running ollama pull qwen2.5vl:7b, and if not installed, only the text check runs.
50%UnverifiedRealist's findings fall into four categories: text_conflict (fields contradict each other), image_mismatch (photos do not match the text), possible_omission (visible content is not stated), and unverified_claim (claims unsupported by materials, not accusations).
50%UnverifiedRealist splits the work into two parts: pure Python code handles precise comparisons (such as Bluetooth version, weight, color, etc.), and the vision model only returns structured observation results without making judgments.
50%UnverifiedBecause the Python comparison is deterministic and involves no sampling or temperature parameters, the confidence of text conflicts (text_conflict) is fixed at 1.0.
50%UnverifiedThe confidence cap for photo-based color mismatches is set at 0.8, because lighting and image processing can change colors; when the model number digits recognized by OCR are close to but different from the listing, the confidence is 0.5, because OCR can misread characters.
50%UnverifiedRealist runs the Qwen2.5-VL-7B vision model through a local Ollama server (default qwen2.5vl:7b), and photos are scaled down and stripped of metadata before being sent only to 127.0.0.1, not flowing externally.
50%UnverifiedRealist's visual audit relies on a three-line VisionBackend abstraction interface, making vision models such as Qwen2.5-VL replaceable components.
50%UnverifiedRealist uses PaddleOCR as an optional OCR layer to extract text such as model numbers and labels from photos and compare them against the listing; both the model and PaddleOCR use the Apache-2.0 license.
50%UnverifiedRealist has tested the deterministic engine, including 10 unit tests covering matching listings producing zero findings, constructed conflicts being caught, unit equivalences not being flagged, specification-supported claims not being flagged, specification-refuted claims being caught, and HTML export escaping malicious input, among others.
50%UnverifiedRealist has not yet tested the real observation accuracy of photo inspection, because it uses handwritten observation results; it also has not yet built an annotated evaluation set with precision, recall, and false positive rates, nor a pure-model baseline.
50%UnverifiedThe author hypothesizes that the hybrid approach produces fewer false positives and more actionable findings than 'asking the LLM whether it looks right', but this hypothesis has not yet been measured.
50%UnverifiedRealist is an open-source product listing auditing tool used to perform quality checks on sellers' existing product listings and discover contradictions or inconsistencies among titles, descriptions, specifications, and photos.
50%UnverifiedRealist's pricing is 5 audits free, then roughly $9 to $29 per month, but the author states that these prices are unverified guesses, not market prices.
50%UnverifiedRealist's next steps include: building an evaluation set of 30-50 listings and publishing the numbers (including baseline comparisons), adding rule families for dimensions/voltage and power/compatibility/materials, and offering bulk auditing and downloadable reports for organizations.
50%