Cleanup and update readme

Author: smoltuna

README.md

@@ -18,7 +18,7 @@ class Settings(BaseSettings):
     together_api_key: str = ""
     openai_api_key: str = ""
 
-    llm_provider: str = "ollama"  # ollama | gemini | groq | together | openai
+    llm_provider: str = "ollama"  # ollama | gemini | groq | together
 
     # Separate provider for translation — defaults to ollama (no rate limits, runs locally).
     # Set TRANSLATION_PROVIDER=groq in .env only if you have a paid Groq account with
@@ -34,10 +34,10 @@ class Settings(BaseSettings):
     # Per-step Ollama model overrides
     grade_model: str = ""                   # e.g. "llama3.2:1b" — falls back to ollama_llm_model
     query_gen_model: str = ""               # e.g. "llama3.2:3b"
-    synth_model: str = ""                   # e.g. "qwen2.5:7b" — recommended for fewer hallucinations
+    synth_model: str = ""                   # e.g. "qwen2.5:7b" for fewer hallucinations
     judge_model: str = ""                   # e.g. "llama3.2:3b"
-    translation_model: str = ""             # e.g. "qwen2.5:7b" — strong multilingual on M4 Metal
-    fact_check_model: str = ""              # e.g. "qwen2.5:7b" — used to compare claim vs web snippet
+    translation_model: str = ""             # e.g. "qwen2.5:7b" for multilingual on Metal
+    fact_check_model: str = ""              # e.g. "qwen2.5:7b" compares claim vs snippet
 
     # Post-synthesis web fact-check for COMPLEX fields (etymology, cultural_info, fun_fact).
     # Runs a targeted DDG search per field, compares the synthesized claim to the top snippet,

backend/config.py

@@ -22,6 +22,3 @@ def configure_logging() -> None:
         logger_factory=structlog.PrintLoggerFactory(),
         cache_logger_on_first_use=True,
     )
-
-
-logger = structlog.get_logger()

backend/log_config.py

@@ -60,9 +60,15 @@ class Flower(Base):
         DateTime(timezone=True), server_default=func.now(), onupdate=func.now()
     )
 
-    raw_sources: Mapped[list["RawSource"]] = relationship(back_populates="flower", cascade="all, delete-orphan")
-    embeddings: Mapped[list["SourceEmbedding"]] = relationship(back_populates="flower", cascade="all, delete-orphan")
-    translations: Mapped[list["Translation"]] = relationship(back_populates="flower", cascade="all, delete-orphan")
+    raw_sources: Mapped[list["RawSource"]] = relationship(
+        back_populates="flower", cascade="all, delete-orphan"
+    )
+    embeddings: Mapped[list["SourceEmbedding"]] = relationship(
+        back_populates="flower", cascade="all, delete-orphan"
+    )
+    translations: Mapped[list["Translation"]] = relationship(
+        back_populates="flower", cascade="all, delete-orphan"
+    )
 
 
 class RawSource(Base):

backend/logging.py

@@ -3,7 +3,7 @@
 
 from pathlib import Path
 
-from database import async_session_factory, get_db
+from database import get_db
 from fastapi import APIRouter, Depends, HTTPException
 from fastapi.responses import FileResponse
 from models import Flower
@@ -109,7 +109,7 @@ async def run_image_pipeline(flower_id: int, db: AsyncSession = Depends(get_db))
     from config import settings
     from services.images.lock_gen import generate_lock_image
     from services.images.processor import process_info_image, process_main_image
-    from services.images.wikimedia import find_images
+    from services.images.search import find_images
 
     flower = await db.get(Flower, flower_id)
     if not flower:

backend/models.py

@@ -282,16 +282,3 @@ def _add(imgs: list[WikimediaImage]) -> None:
 
     return candidates
 
-
-# ---------------------------------------------------------------------------
-# Legacy API — kept for backwards compatibility with existing callers
-# ---------------------------------------------------------------------------
-
-async def find_images(latin_name: str) -> ImagePair:
-    """Return the best (info, blossom) pair from Wikimedia Commons.
-
-    DEPRECATED: Use search.find_images() instead for multi-source results.
-    Kept for backwards compatibility.
-    """
-    from services.images.search import find_images as _find
-    return await _find(latin_name)

backend/routers/flowers.py

@@ -49,9 +49,6 @@ def get_provider(
     elif name == "together":
         from services.llm.together import TogetherProvider
         return TogetherProvider()
-    elif name == "openai":
-        from services.llm.openai import OpenAIProvider
-        return OpenAIProvider()
     else:
         from services.llm.ollama import OllamaProvider
         return OllamaProvider(model_override=step_model)

backend/services/images/wikimedia.py

@@ -164,6 +164,8 @@ def on_end(self, span: ReadableSpan) -> None:
             import mlflow
             if not mlflow.active_run():
                 return
+            if span.end_time is None or span.start_time is None:
+                return
             duration_s = (span.end_time - span.start_time) / 1_000_000_000.0
             name = span.name
             attrs = span.attributes or {}
@@ -181,7 +183,7 @@ def on_end(self, span: ReadableSpan) -> None:
             metrics: dict[str, float] = {f"trace_{name}_s": duration_s}
             for key in ("tokens_used", "llm_calls", "api_calls", "chunks_in", "chunks_out"):
                 v = attrs.get(key)
-                if v:
+                if isinstance(v, (int, float)):
                     # Normalize "tokens_used" → "tokens" for backward-compatible metric names.
                     metric_key = "tokens" if key == "tokens_used" else key
                     metrics[f"trace_{name}_{metric_key}"] = float(v)
@@ -216,6 +218,8 @@ def on_start(self, span: Span, parent_context: Context | None = None) -> None:
 
     def on_end(self, span: ReadableSpan) -> None:
         trace_id = span.context.trace_id
+        if span.end_time is None or span.start_time is None:
+            return
         duration_s = (span.end_time - span.start_time) / 1_000_000_000.0
         attrs = span.attributes or {}
         entry = self.flowers[trace_id]
@@ -226,8 +230,10 @@ def on_end(self, span: ReadableSpan) -> None:
             entry["total_s"] = duration_s
             return
 
-        tokens = int(attrs.get("tokens_used", 0) or 0)
-        calls = int(attrs.get("llm_calls", 0) or 0)
+        raw_tokens = attrs.get("tokens_used", 0) or 0
+        raw_calls = attrs.get("llm_calls", 0) or 0
+        tokens = int(raw_tokens) if isinstance(raw_tokens, (int, float)) else 0
+        calls = int(raw_calls) if isinstance(raw_calls, (int, float)) else 0
         entry["steps"][span.name] = {
             "duration_s": duration_s,
             "tokens": tokens,

backend/services/llm/provider.py

@@ -74,26 +74,6 @@ async def embed_and_store(
     return results
 
 
-async def embed_all_sources(
-    flower_id: int,
-    embed_provider: EmbeddingProvider,
-    session: AsyncSession,
-) -> list[SourceEmbedding]:
-    """Embed all raw sources for a flower that don't yet have embeddings."""
-    result = await session.execute(
-        select(RawSource).where(RawSource.flower_id == flower_id)
-    )
-    sources = result.scalars().all()
-
-    all_embeddings: list[SourceEmbedding] = []
-    for src in sources:
-        if not src.raw_content and not src.parsed_content:
-            continue
-        embs = await embed_and_store(flower_id, src, embed_provider, session)
-        all_embeddings.extend(embs)
-    return all_embeddings
-
-
 def _build_chunk_text(raw_source: RawSource) -> str:
     """Concatenate available text from a raw source into a single string."""
     parts: list[str] = []