Switch default Qwen model and collapse Q&A snippets

README.md

@@ -10,7 +10,7 @@ app_file: app.py
 pinned: false
 short_description: Simplify complex legislation that affects you!
 models:
-  - Qwen/Qwen3-32B
+  - Qwen/Qwen3-14B
 ---
 
 # Legislation Explainer
@@ -29,13 +29,9 @@ It is created for the Hugging Face Build Small Hackathon under the `Backyard AI`
 
 - Track: `Backyard AI`
 - Real user: Ghanaian citizens and digital-policy stakeholders who need a clearer view of a bill's practical effects.
-- Small-model constraint: default provider is `Qwen/Qwen3-32B:cheapest` through the Hugging Face router, staying within the hackathon's `<= 32B` model cap. I could go lower than 32B, but since this deals with legal documents I want the best cognition possible.
+- Small-model constraint: default provider is `Qwen/Qwen3-14B:cheapest` through the Hugging Face router, staying within the hackathon's `<= 32B` model cap while giving a better speed/cost tradeoff for this app.
 - Required surface: Gradio app, ready for Hugging Face Spaces through `app.py`.
-<<<<<<< HEAD
-- Submission assets still needed: Space link, short demo video, and social post.
-=======
 - GitHub repo: https://github.com/KayO-GH/legislation-explainer
->>>>>>> hf
 
 ## What It Does
 
@@ -52,7 +48,7 @@ It is created for the Hugging Face Build Small Hackathon under the `Backyard AI`
 
 ## Model And Provider Notes
 
-The hackathon-safe default is Qwen3 32B through the Hugging Face router. Bring-your-own provider support is a proposed future expansion and is currently commented out so the hackathon build stays focused on one documented `<= 32B` model path.
+The hackathon-safe default is Qwen3 14B through the Hugging Face router. Bring-your-own provider support is a proposed future expansion and is currently commented out so the hackathon build stays focused on one documented `<= 32B` model path.
 
 For future expansion, we will have a bring your own provider setting allowing users to connect to other models eg. from OpenAI, ANthropic, etc. if they so wish.
 

app.py

@@ -383,6 +383,25 @@ def _format_analysis(analysis: AnalysisResult | None) -> str:
     )
 
 
+def _render_supporting_snippets(citations: list[dict[str, Any]]) -> str:
+    if not citations:
+        return ""
+
+    items = []
+    for item in citations:
+        ref_id = html.escape(str(item["ref_id"]))
+        snippet = html.escape(item["snippet"])
+        items.append(f"<li><strong>[ref{ref_id}]</strong> {snippet}</li>")
+
+    count = len(citations)
+    label = "snippet" if count == 1 else "snippets"
+    return (
+        f"<details><summary>Supporting {label} ({count})</summary>"
+        f"<ul>{''.join(items)}</ul>"
+        "</details>"
+    )
+
+
 def _format_chat_entry(answer_text: str, citations: list[dict[str, Any]], *, provenance: str = "analysis_based") -> str:
     prefix = ""
     if provenance == "analysis_based":
@@ -392,8 +411,7 @@ def _format_chat_entry(answer_text: str, citations: list[dict[str, Any]], *, pro
 
     if not citations:
         return prefix + answer_text
-    citations_text = "\n".join(f"- [ref{item['ref_id']}] {item['snippet']}" for item in citations)
-    return f"{prefix}{answer_text}\n\nSupporting snippets:\n{citations_text}"
+    return f"{prefix}{answer_text}\n\n{_render_supporting_snippets(citations)}"
 
 
 def _stream_chat_entry(
@@ -403,13 +421,21 @@ def _stream_chat_entry(
     *,
     provenance: str = "analysis_based",
 ):
-    formatted_answer = _format_chat_entry(answer_text, citations, provenance=provenance)
+    prefix = ""
+    if provenance == "analysis_based":
+        prefix = "_Based on the summary and analysis._\n\n"
+    elif provenance == "full_document":
+        prefix = "_Deeper full-document answer._\n\n"
+
+    formatted_answer = prefix + answer_text
     streamed_answer = ""
     for character in formatted_answer:
         streamed_answer += character
         yield base_history + [{"role": "assistant", "content": streamed_answer}]
         if CHAT_STREAM_DELAY_SECONDS:
             time.sleep(CHAT_STREAM_DELAY_SECONDS)
+    if citations:
+        yield base_history + [{"role": "assistant", "content": _format_chat_entry(answer_text, citations, provenance=provenance)}]
 
 
 def _deeper_answer_updates(visible: bool, label: str = "Run deeper full-document answer") -> gr.update:

assets/example_bills/cybersecurity-amendment-bill-2025/metadata.json

@@ -1,10 +1,10 @@
 {
   "generated_at": "2026-06-01T11:48:29.371678+00:00",
   "provider": "qwen",
-  "model": "Qwen/Qwen3-32B:cheapest",
+  "model": "Qwen/Qwen3-14B:cheapest",
   "source_url": "https://www.csa.gov.gh/resources/Cybersecurity%20%28Amendment%29%20Draft%20Bill%202025%20final%2015102025.pdf",
   "document_hash": "d879a4473b3631906b7c4e965195d53f4693a722ce7644e187ab76fca33ef763",
   "chunk_size": 350,
   "chunk_overlap": 60,
   "embedding_model": "sentence-transformers/all-MiniLM-L6-v2"
-}
+}

config.py

@@ -34,7 +34,7 @@ warnings.filterwarnings(
 
 SUPPORTED_PROVIDERS = ["qwen", "openai", "anthropic", "gemini", "cohere"]
 DEFAULT_PROVIDER: str = "qwen"
-DEFAULT_QWEN_MODEL = "Qwen/Qwen3-32B:cheapest"
+DEFAULT_QWEN_MODEL = "Qwen/Qwen3-14B:cheapest"
 DEFAULT_EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
 OPENAI_REASONING_EFFORT = "medium"
 ANTHROPIC_THINKING_BUDGET = 2048
@@ -90,7 +90,7 @@ PROVIDER_METADATA: list[ProviderConfig] = [
     ProviderConfig(
         name="qwen",
         key_prefix=None,
-        display_name="Qwen3 32B",
+        display_name="Qwen3 14B",
         instructions=(
             "Use your Hugging Face token for the router-backed Qwen model. Leave blank to use HF_TOKEN from .env if configured."
         ),

tests/test_app.py

@@ -0,0 +1,26 @@
+from app import _format_chat_entry, _stream_chat_entry
+
+
+def test_format_chat_entry_wraps_supporting_snippets_in_details() -> None:
+    result = _format_chat_entry(
+        "Answer text",
+        [{"ref_id": 1, "snippet": "Quoted <clause> & detail"}],
+    )
+
+    assert "Answer text" in result
+    assert "<details><summary>Supporting snippet (1)</summary>" in result
+    assert "[ref1]" in result
+    assert "Quoted &lt;clause&gt; &amp; detail" in result
+
+
+def test_stream_chat_entry_appends_collapsible_snippets_after_text_stream() -> None:
+    frames = list(
+        _stream_chat_entry(
+            [{"role": "user", "content": "Question"}],
+            "Answer text",
+            [{"ref_id": 1, "snippet": "Snippet text"}],
+        )
+    )
+
+    assert frames[-2][-1]["content"] == "_Based on the summary and analysis._\n\nAnswer text"
+    assert "<details><summary>Supporting snippet (1)</summary>" in frames[-1][-1]["content"]