Deploy ae593cc from hugging_face

app.py

@@ -29,7 +29,9 @@ ICON_DIR = BASE_DIR / "src" / "temp"
 ICON_DIR.mkdir(parents=True, exist_ok=True)
 
 LLM_CHOICES = ["hf", "llama", "gemini", "bedrock"]
-TRANSLATION_CHOICES = ["google", "gemini", "nllb"]
+# Translation engine: 'Automatic' routes by target language; 'Gemini' uses Gemini
+# for everything (except KSL gloss, which only NLLB can produce).
+TRANSLATION_ENGINES = ["Automatic", "Gemini"]
 
 LANGUAGES = {
     "None (keep original)": "",
@@ -41,10 +43,22 @@ LANGUAGES = {
     "German": "de",
 }
 
-# The NLLB fine-tune (EzekielMW/LuoKslGloss) was only trained on these targets,
-# and KSL gloss is not something the general-purpose translators can produce.
-NLLB_LANGUAGES = {"sw", "swa", "luo", "ksl", "en", "eng"}
-NLLB_ONLY_LANGUAGES = {"ksl"}
+# Per-language routing (matches the Phase II report): the local NLLB-200 fine-tune
+# handles Swahili and Luo; Google Translate handles French/Spanish/German. KSL
+# gloss is NLLB-only.
+_NLLB_ROUTED = {"sw", "swa", "luo"}
+
+
+def resolve_translation_provider(engine: str, code: str) -> str:
+    """Choose the translation backend for a target language code."""
+    code = (code or "").lower()
+    if not code:
+        return "google"  # no translation requested; provider is unused
+    if code == "ksl":
+        return "nllb"  # only NLLB produces sign-language gloss
+    if engine == "Gemini":
+        return "gemini"  # Gemini alternative for all other languages
+    return "nllb" if code in _NLLB_ROUTED else "google"
 
 VOICES = ["af_heart", "af_bella", "am_michael", "bf_lily", "bm_george"]
 
@@ -85,22 +99,6 @@ def get_tts():
     return _tts
 
 
-def check_language(translation_provider: str, code: str) -> None:
-    """Reject provider/language pairs that would silently return untranslated text."""
-    if not code:
-        return
-    if translation_provider == "nllb" and code not in NLLB_LANGUAGES:
-        raise gr.Error(
-            "The NLLB model only translates into Swahili, Luo and KSL gloss. "
-            "Pick one of those, or switch Translation to 'google'."
-        )
-    if code in NLLB_ONLY_LANGUAGES and translation_provider != "nllb":
-        raise gr.Error(
-            "Kenyan Sign Language gloss is only produced by the 'nllb' "
-            "translation provider. Switch Translation to 'nllb'."
-        )
-
-
 def read_upload(file_path: str | None) -> str:
     """Extract plain text from an uploaded .txt/.md/.docx/.pdf document."""
     if not file_path:
@@ -221,14 +219,14 @@ def render_sentences(state: dict) -> str:
     return f'<h2 style="margin-bottom:14px;">{title}</h2>' + "".join(rows)
 
 
-def step_simplify(text, llm_provider, translation_provider, language_label):
+def step_simplify(text, llm_provider, translation_engine, language_label):
     if not text or not text.strip():
         raise gr.Error("Add some text — upload a document or paste it above.")
 
     target_language = LANGUAGES.get(language_label, "")
-    check_language(translation_provider, target_language)
+    provider = resolve_translation_provider(translation_engine, target_language)
 
-    controller = get_controller(llm_provider, translation_provider)
+    controller = get_controller(llm_provider, provider)
     result = controller.simplify_text(text, target_language=target_language or None)
     if "error" in result:
         raise gr.Error(f"The model did not return valid JSON: {result['error']}")
@@ -238,6 +236,7 @@ def step_simplify(text, llm_provider, translation_provider, language_label):
         "sentences": result.get("simplified_sentences", []),
         "original": text,
         "language": target_language,
+        "translation_engine": translation_engine,
         "request_id": None,
         "feedback": "",
     }
@@ -256,12 +255,13 @@ def step_simplify(text, llm_provider, translation_provider, language_label):
     )
 
 
-def step_validate(state, rows, llm_provider, translation_provider):
+def step_validate(state, rows, llm_provider, translation_engine):
     state = sync_table(state, rows)
     if not state.get("sentences"):
         raise gr.Error("Run Simplify first.")
 
-    controller = get_controller(llm_provider, translation_provider)
+    provider = resolve_translation_provider(translation_engine, state.get("language", ""))
+    controller = get_controller(llm_provider, provider)
     result = controller.validate_text(state["original"], state["sentences"])
     if "error" in result:
         raise gr.Error(f"Validation did not return valid JSON: {result['error']}")
@@ -282,12 +282,13 @@ def step_validate(state, rows, llm_provider, translation_provider):
     return state, banner + detail
 
 
-def step_revise(state, rows, llm_provider, translation_provider):
+def step_revise(state, rows, llm_provider, translation_engine):
     state = sync_table(state, rows)
     if not state.get("feedback"):
         raise gr.Error("Run Validate first — Revise consumes its feedback.")
 
-    controller = get_controller(llm_provider, translation_provider)
+    provider = resolve_translation_provider(translation_engine, state.get("language", ""))
+    controller = get_controller(llm_provider, provider)
     result = controller.revise_text(
         original_text=state["original"],
         easy_read_sentences=state["sentences"],
@@ -323,12 +324,13 @@ def _generate_button_update(sentences: list[dict]):
     return gr.update(visible=True, value=label)
 
 
-def step_find_symbols(state, rows, llm_provider, translation_provider, symbolset):
+def step_find_symbols(state, rows, llm_provider, translation_engine, symbolset):
     state = sync_table(state, rows)
     if not state.get("sentences"):
         raise gr.Error("Run Simplify first.")
 
-    controller = get_controller(llm_provider, translation_provider)
+    provider = resolve_translation_provider(translation_engine, state.get("language", ""))
+    controller = get_controller(llm_provider, provider)
     result = controller.search_symbols(state["sentences"], symbolset=symbolset)
     state["request_id"] = result["request_id"]
     for sentence, found in zip(state["sentences"], result["results"]):
@@ -338,7 +340,7 @@ def step_find_symbols(state, rows, llm_provider, translation_provider, symbolset
     return state, render_sentences(state), _generate_button_update(state["sentences"])
 
 
-def step_generate_missing(state, rows, llm_provider, translation_provider):
+def step_generate_missing(state, rows, llm_provider, translation_engine):
     state = sync_table(state, rows)
     if not state.get("sentences"):
         raise gr.Error("Run Simplify first.")
@@ -348,7 +350,8 @@ def step_generate_missing(state, rows, llm_provider, translation_provider):
         gr.Info("Every sentence already has a symbol — nothing to generate.")
         return state, render_sentences(state), _generate_button_update(state["sentences"])
 
-    controller = get_controller(llm_provider, translation_provider)
+    provider = resolve_translation_provider(translation_engine, state.get("language", ""))
+    controller = get_controller(llm_provider, provider)
     # Generation only — no symbol lookup (Find symbols is the separate search
     # step). Uses the current, table-synced prompt for each missing sentence.
     controller.generate_icons(misses, use_global_symbols=False)
@@ -434,12 +437,12 @@ with gr.Blocks(title="GenAI for Easy Read") as demo:
             info="'hf' runs Llama via Hugging Face Inference (needs HF_TOKEN). "
             "'llama'/'bedrock' use AWS; 'gemini' uses a Google key.",
         )
-        translation_provider = gr.Dropdown(
-            TRANSLATION_CHOICES,
-            value="google",
-            label="Translation",
-            info="'nllb' loads a 1.3B model on first use — the only option for "
-            "Luo and KSL gloss.",
+        translation_engine = gr.Dropdown(
+            TRANSLATION_ENGINES,
+            value="Automatic",
+            label="Translation engine",
+            info="Automatic routes by language: NLLB for Swahili/Luo/KSL, Google "
+            "for French/Spanish/German. Gemini translates all languages.",
         )
         language = gr.Dropdown(
             list(LANGUAGES), value="None (keep original)", label="Translate into"
@@ -516,31 +519,31 @@ with gr.Blocks(title="GenAI for Easy Read") as demo:
     upload.upload(read_upload, inputs=[upload], outputs=[text_input])
     simplify_btn.click(
         step_simplify,
-        inputs=[text_input, llm_provider, translation_provider, language],
+        inputs=[text_input, llm_provider, translation_engine, language],
         outputs=[state, table, feedback_out, review_group, preview, empty_hint,
                  export_group, generate_missing_btn],
         scroll_to_output=True,
     )
     validate_btn.click(
         step_validate,
-        inputs=[state, table, llm_provider, translation_provider],
+        inputs=[state, table, llm_provider, translation_engine],
         outputs=[state, feedback_out],
     )
     revise_btn.click(
         step_revise,
-        inputs=[state, table, llm_provider, translation_provider],
+        inputs=[state, table, llm_provider, translation_engine],
         outputs=[state, table],
     )
     illustrate_btn.click(
         step_find_symbols,
-        inputs=[state, table, llm_provider, translation_provider, symbolset],
+        inputs=[state, table, llm_provider, translation_engine, symbolset],
         outputs=[state, preview, generate_missing_btn],
         # Auto-scroll to the results so they aren't lost below the fold.
         scroll_to_output=True,
     )
     generate_missing_btn.click(
         step_generate_missing,
-        inputs=[state, table, llm_provider, translation_provider],
+        inputs=[state, table, llm_provider, translation_engine],
         outputs=[state, preview, generate_missing_btn],
         scroll_to_output=True,
     )