Marconcini commited on
Commit
a5f0ff2
·
verified ·
1 Parent(s): 836740a

README: link the cited RAG assistant and its results

Browse files
Files changed (1) hide show
  1. README.md +17 -1
README.md CHANGED
@@ -24,4 +24,20 @@ right. A research demo, not a diagnostic tool.
24
  Code, experiments and limitations: [github.com/marconcini/leafscan](https://github.com/marconcini/leafscan) ·
25
  Model: [Marconcini/leafscan-convnext-tiny](https://huggingface.co/Marconcini/leafscan-convnext-tiny)
26
 
27
- The Python/Gradio version of the demo (with Grad-CAM) runs locally: `python -m leafscan.app`.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
24
  Code, experiments and limitations: [github.com/marconcini/leafscan](https://github.com/marconcini/leafscan) ·
25
  Model: [Marconcini/leafscan-convnext-tiny](https://huggingface.co/Marconcini/leafscan-convnext-tiny)
26
 
27
+ ## Beyond this page: a cited assistant (local only)
28
+
29
+ The full project adds a question-answering assistant after the diagnosis. It retrieves passages
30
+ from 28 Wikipedia articles, answers with a local LLM (Qwen3-4B) citing each source, says
31
+ "I don't know" when the sources do not answer, and flags with ⚠️ the sentences an NLI model
32
+ finds unsupported. Its models are ~10 GB, so it runs on your machine, not in this page:
33
+
34
+ ```bash
35
+ pip install -e ".[app,rag]"
36
+ LEAFSCAN_ASSISTANT=1 python -m leafscan.app # photo → diagnosis → cited answers, Grad-CAM
37
+ ```
38
+
39
+ What the evaluation showed: the assistant abstains on 86% of unanswerable test questions and
40
+ 78% of its cited sentences are supported by their source; letting the diagnosis filter the
41
+ search would lose 16 points of recall, because a wrong diagnosis removes the answer. Details
42
+ in the [README](https://github.com/marconcini/leafscan#phase-2--rag-assistant) and the
43
+ [research log](https://github.com/marconcini/leafscan/blob/main/docs/research-log.md).