feat: Change out storage from local to Bunny (via env param)

api/app/config.py

@@ -15,10 +15,16 @@ class Settings(BaseSettings):
     scaleway_tem_project_id: str = ""
     scaleway_tem_from_address: str = ""
     scaleway_tem_region: str = "fr-par"
-    storage_endpoint_url: str
+    storage_provider: str = "local"  # or 'bunny'
-    storage_access_key: str
+    storage_endpoint_url: str = ""
-    storage_secret_key: str
+    storage_access_key: str = ""
+    storage_secret_key: str = ""
     storage_bucket: str = "langlearn"
+    bunny_zone: str = "languagelearningapp"
+    bunny_api_key: str = ""
+    bunny_cdn_base_url: str = ""
+    bunny_token_auth_key: str = ""
+    bunny_storage_endpoint: str = "https://storage.bunnycdn.com"
     stub_generation: bool = False
 
     model_config = {"env_file": ".env"}

api/app/domain/services/adventure_service.py

@@ -24,7 +24,7 @@ from ...outbound.postgres.repositories.adventure_repository import (
     PostgresAdventureRepository,
 )
 from ...outbound.spacy.spacy_client import SpacyClient
-from ...storage import upload_audio
+from ...outbound.storage_client import get_storage_client
 from ..models.adventure import (
     Adventure,
     AdventureEntry,
@@ -251,7 +251,7 @@ class AdventureService:
                 for sent_idx, target_sent in enumerate(target_nlp["sentences"]):
                     t0 = time.monotonic()
                     translated_sentence = await self.deepl_client.translate(
-                        target_sent["text"], adventure.source_language
+                        target_sent["text"], adventure.source_language, paragraph_text
                     )
                     timing_translations += time.monotonic() - t0
 
@@ -315,7 +315,7 @@ class AdventureService:
             # ── File upload ───────────────────────────────────────────────────
             t0 = time.monotonic()
             audio_key = f"adventure-audio/{entry_id}.wav"
-            upload_audio(audio_key, wav_bytes)
+            get_storage_client().upload(audio_key, wav_bytes)
             timing_file_uploading = time.monotonic() - t0
 
             await self.audio_repo.create(

api/app/domain/services/summarise_service.py

@@ -14,7 +14,7 @@ from ...outbound.deepgram.deepgram_client import LocalDeepgramClient
 from ...outbound.deepl.deepl_client import DeepLClient
 from ...outbound.gemini.gemini_client import GeminiClient
 from ...outbound.spacy.spacy_client import SpacyClient
-from ...storage import upload_audio
+from ...outbound.storage_client import get_storage_client
 from ...languages import SUPPORTED_LANGUAGES
 
 
@@ -137,7 +137,7 @@ class SummariseService:
             voice = self.gemini_client.get_voice_by_language(target_language)
             wav_bytes = await self.gemini_client.generate_audio(generated_text, voice)
             audio_key = f"audio/{job_id}.wav"
-            upload_audio(audio_key, wav_bytes)
+            get_storage_client().upload(audio_key, wav_bytes)
 
             transcript = await self.deepgram_client.transcribe_bytes(wav_bytes, target_language)
 

api/app/main.py

@@ -8,13 +8,13 @@ from .routers.api import jobs
 from .routers import media as media_router
 from .routers.api.main import api_router
 from .routers.bff.main import bff_router
-from .storage import ensure_bucket_exists
+from .outbound.storage_factory import init_storage
 from . import worker
 
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    ensure_bucket_exists()
+    init_storage()
     worker_task = asyncio.create_task(worker.worker_loop())
     yield
     worker_task.cancel()

api/app/outbound/bunny/bunny_client.py

@@ -0,0 +1,77 @@
+import base64
+import hashlib
+import time
+import urllib.error
+import urllib.request
+
+_SIGNED_URL_EXPIRY_SECONDS = 3600
+
+
+class BunnyClient:
+    def __init__(
+        self,
+        zone: str,
+        api_key: str,
+        cdn_base_url: str,
+        token_auth_key: str,
+        storage_endpoint: str = "https://storage.bunnycdn.com",
+    ) -> None:
+        self._zone = zone
+        self._api_key = api_key
+        self._cdn_base_url = cdn_base_url.rstrip("/")
+        self._token_auth_key = token_auth_key
+        self._storage_endpoint = storage_endpoint.rstrip("/")
+
+    def _storage_url(self, path: str) -> str:
+        return f"{self._storage_endpoint}/{self._zone}/{path.lstrip('/')}"
+
+    def upload(self, path: str, data: bytes) -> bool:
+        req = urllib.request.Request(
+            self._storage_url(path),
+            data=data,
+            method="PUT",
+            headers={
+                "AccessKey": self._api_key,
+                "Content-Type": "audio/wav",
+            },
+        )
+        try:
+            with urllib.request.urlopen(req) as resp:
+                return resp.status == 201
+        except urllib.error.HTTPError:
+            return False
+
+    def get_url(self, path: str) -> str:
+        url_path = f"/{path.lstrip('/')}"
+        expiration = int(time.time()) + _SIGNED_URL_EXPIRY_SECONDS
+        digest = hashlib.sha256(
+            (self._token_auth_key + url_path + str(expiration)).encode()
+        ).digest()
+        token = (
+            base64.b64encode(digest)
+            .decode()
+            .replace("+", "-")
+            .replace("/", "_")
+            .replace("=", "")
+        )
+        return f"{self._cdn_base_url}{url_path}?token={token}&expires={expiration}"
+
+    def get_public_url(self, path: str) -> str:
+        return f"{self._cdn_base_url}/{path.lstrip('/')}"
+
+    def delete(self, path: str) -> bool:
+        req = urllib.request.Request(
+            self._storage_url(path),
+            method="DELETE",
+            headers={"AccessKey": self._api_key},
+        )
+        try:
+            with urllib.request.urlopen(req) as resp:
+                return resp.status == 200
+        except urllib.error.HTTPError:
+            return False
+
+    def download(self, path: str) -> tuple[bytes, str]:
+        raise NotImplementedError(
+            "Direct download not available with Bunny — use get_url() to obtain a signed CDN URL"
+        )

api/app/outbound/gemini/gemini_client.py

@@ -1,9 +1,19 @@
 import asyncio
+import io
+import wave
 
 from google import genai
 from google.genai import types as genai_types
 
-from ...storage import pcm_to_wav
+
+def _pcm_to_wav(pcm_data: bytes, sample_rate: int = 24000) -> bytes:
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as wf:
+        wf.setnchannels(1)
+        wf.setsampwidth(2)
+        wf.setframerate(sample_rate)
+        wf.writeframes(pcm_data)
+    return buf.getvalue()
 
 VOICE_BY_LANGUAGE: dict[str, str] = {
     "fr": "Kore",
@@ -47,6 +57,6 @@ class GeminiClient():
                 ),
             )
             pcm_data = response.candidates[0].content.parts[0].inline_data.data
-            return pcm_to_wav(pcm_data)
+            return _pcm_to_wav(pcm_data)
 
         return await asyncio.to_thread(_call)

api/app/outbound/minio/minio_client.py

@@ -0,0 +1,70 @@
+import boto3
+from botocore.exceptions import ClientError
+
+
+class MinioClient:
+    def __init__(
+        self,
+        endpoint_url: str,
+        access_key: str,
+        secret_key: str,
+        bucket: str,
+        api_base_url: str,
+    ) -> None:
+        self._endpoint_url = endpoint_url
+        self._access_key = access_key
+        self._secret_key = secret_key
+        self._bucket = bucket
+        self._api_base_url = api_base_url.rstrip("/")
+
+    def _s3(self):
+        return boto3.client(
+            "s3",
+            endpoint_url=self._endpoint_url,
+            aws_access_key_id=self._access_key,
+            aws_secret_access_key=self._secret_key,
+        )
+
+    def ensure_bucket_exists(self) -> None:
+        client = self._s3()
+        try:
+            client.head_bucket(Bucket=self._bucket)
+        except ClientError as e:
+            if e.response["Error"]["Code"] in ("404", "NoSuchBucket"):
+                client.create_bucket(Bucket=self._bucket)
+            else:
+                raise
+
+    def upload(self, path: str, data: bytes) -> bool:
+        try:
+            self._s3().put_object(
+                Bucket=self._bucket,
+                Key=path,
+                Body=data,
+                ContentType="audio/wav",
+            )
+            return True
+        except ClientError:
+            return False
+
+    def get_url(self, path: str) -> str:
+        return f"{self._api_base_url}/media/{path}"
+
+    def get_public_url(self, path: str) -> str:
+        return f"{self._api_base_url}/media/{path}"
+
+    def delete(self, path: str) -> bool:
+        try:
+            self._s3().delete_object(Bucket=self._bucket, Key=path)
+            return True
+        except ClientError:
+            return False
+
+    def download(self, path: str) -> tuple[bytes, str]:
+        try:
+            response = self._s3().get_object(Bucket=self._bucket, Key=path)
+            return response["Body"].read(), response.get("ContentType", "audio/wav")
+        except ClientError as e:
+            if e.response["Error"]["Code"] in ("NoSuchKey", "404"):
+                raise FileNotFoundError(path)
+            raise

api/app/outbound/storage_client.py

@@ -0,0 +1,21 @@
+from typing import Protocol
+
+_client: "StorageClient | None" = None
+
+
+class StorageClient(Protocol):
+    def upload(self, path: str, data: bytes) -> bool: ...
+    def get_url(self, path: str) -> str: ...
+    def get_public_url(self, path: str) -> str: ...
+    def delete(self, path: str) -> bool: ...
+    def download(self, path: str) -> tuple[bytes, str]: ...
+
+
+def get_storage_client() -> "StorageClient":
+    assert _client is not None, "Storage client not initialised — call init_storage() at startup"
+    return _client
+
+
+def _set_storage_client(c: "StorageClient") -> None:
+    global _client
+    _client = c

api/app/outbound/storage_factory.py

@@ -0,0 +1,27 @@
+from ..config import settings
+from .storage_client import StorageClient, _set_storage_client
+from .minio.minio_client import MinioClient
+from .bunny.bunny_client import BunnyClient
+
+
+def init_storage() -> None:
+    client: StorageClient
+    if settings.storage_provider == "bunny":
+        client = BunnyClient(
+            zone=settings.bunny_zone,
+            api_key=settings.bunny_api_key,
+            cdn_base_url=settings.bunny_cdn_base_url,
+            token_auth_key=settings.bunny_token_auth_key,
+            storage_endpoint=settings.bunny_storage_endpoint,
+        )
+    else:
+        minio = MinioClient(
+            endpoint_url=settings.storage_endpoint_url,
+            access_key=settings.storage_access_key,
+            secret_key=settings.storage_secret_key,
+            bucket=settings.storage_bucket,
+            api_base_url=settings.api_base_url,
+        )
+        minio.ensure_bucket_exists()
+        client = minio
+    _set_storage_client(client)

api/app/routers/api/jobs.py

@@ -12,7 +12,7 @@ from ...outbound.postgres.repositories import summarise_job_repository
 from ...outbound.postgres.repositories.translated_article_repository import TranslatedArticleRepository
 from ...outbound.postgres.entities.translated_article_entity import TranslatedArticleEntity
 from ...outbound.gemini.gemini_client import GeminiClient
-from ...storage import upload_audio
+from ...outbound.storage_client import get_storage_client
 from ...config import settings
 from ... import worker
 
@@ -92,7 +92,7 @@ async def _run_regenerate_audio(job_id: uuid.UUID) -> None:
             voice = gemini_client.get_voice_by_language(article_entity.target_language)
             wav_bytes = await gemini_client.generate_audio(article_entity.target_body, voice)
             audio_key = f"audio/{job_id}.wav"
-            upload_audio(audio_key, wav_bytes)
+            get_storage_client().upload(audio_key, wav_bytes)
 
             await article_repo.update_audio(
                 article_entity.id,

api/app/routers/bff/adventure.py

@@ -5,8 +5,8 @@ from pydantic import BaseModel
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from ...auth import verify_token
-from ...config import settings
 from ...outbound.postgres.database import get_db
+from ...outbound.storage_client import get_storage_client
 from ...outbound.postgres.repositories.adventure_repository import (
     PostgresAdventureEntryAudioRepository,
     PostgresAdventureEntryChoiceRepository,
@@ -61,7 +61,7 @@ class AdventureDetailResponse(BaseModel):
 def _audio_url(key: str | None) -> str | None:
     if key is None:
         return None
-    return f"{settings.api_base_url}/media/{key}"
+    return get_storage_client().get_url(key)
 
 
 @router.get("/{adventure_id}", response_model=AdventureDetailResponse, status_code=200)

api/app/routers/bff/articles.py

@@ -6,8 +6,8 @@ from pydantic import BaseModel
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from ...auth import verify_token
-from ...config import settings
 from ...outbound.postgres.database import get_db
+from ...outbound.storage_client import get_storage_client
 from ...outbound.postgres.repositories.translated_article_repository import TranslatedArticleRepository
 
 router = APIRouter(prefix="/articles", tags=["bff", "articles"])
@@ -46,7 +46,7 @@ class ArticleDetail(BaseModel):
 def _audio_url(key: str | None) -> str | None:
     if key is None:
         return None
-    return f"{settings.api_base_url}/media/{key}"
+    return get_storage_client().get_url(key)
 
 
 @router.get("", response_model=ArticleListResponse, status_code=200)

api/app/routers/media.py

@@ -3,12 +3,11 @@ import uuid
 from fastapi import APIRouter, Depends, HTTPException
 from fastapi.responses import Response
 from sqlalchemy.ext.asyncio import AsyncSession
-from botocore.exceptions import ClientError
 
 from ..outbound.postgres.database import get_db
 from ..outbound.postgres.repositories.translated_article_repository import TranslatedArticleRepository
 from ..outbound.postgres.repositories.adventure_repository import PostgresAdventureEntryAudioRepository
-from ..storage import download_audio
+from ..outbound.storage_client import get_storage_client
 
 router = APIRouter(prefix="/media", tags=["media"])
 
@@ -23,21 +22,23 @@ async def get_adventure_audio_file(
     except ValueError:
         raise HTTPException(status_code=400, detail="Invalid file ID")
 
-    print(f"Looking for adventure audio with entry ID: {eid}")
-    
     adventure_audio = await PostgresAdventureEntryAudioRepository(db).get_for_entry(entry_id=eid, component_type="story_text")
 
     if adventure_audio is None:
         raise HTTPException(status_code=404, detail="File not found")
+
     try:
-        audio_bytes, content_type = download_audio("adventure-audio/" + filename)
+        audio_bytes, content_type = get_storage_client().download("adventure-audio/" + filename)
-    except ClientError as e:
+    except FileNotFoundError:
-        if e.response["Error"]["Code"] in ("NoSuchKey", "404"):
+        raise HTTPException(status_code=404, detail="File not found")
-            raise HTTPException(status_code=404, detail="File not found")
+    except NotImplementedError:
+        raise HTTPException(status_code=501, detail="Media proxy not available with current storage provider")
+    except Exception:
         raise HTTPException(status_code=500, detail="Storage error")
 
     return Response(content=audio_bytes, media_type=content_type)
 
+
 @router.get("/{filename:path}")
 async def get_media_file(
     filename: str,
@@ -49,11 +50,12 @@ async def get_media_file(
         raise HTTPException(status_code=404, detail="File not found")
 
     try:
-        audio_bytes, content_type = download_audio(filename)
+        audio_bytes, content_type = get_storage_client().download(filename)
-    except ClientError as e:
+    except FileNotFoundError:
-        if e.response["Error"]["Code"] in ("NoSuchKey", "404"):
+        raise HTTPException(status_code=404, detail="File not found")
-            raise HTTPException(status_code=404, detail="File not found")
+    except NotImplementedError:
+        raise HTTPException(status_code=501, detail="Media proxy not available with current storage provider")
+    except Exception:
         raise HTTPException(status_code=500, detail="Storage error")
 
     return Response(content=audio_bytes, media_type=content_type)
-

api/app/storage.py

@@ -1,56 +0,0 @@
-import io
-import wave
-
-import boto3
-from botocore.exceptions import ClientError
-
-from .config import settings
-
-
-def get_s3_client():
-    return boto3.client(
-        "s3",
-        endpoint_url=settings.storage_endpoint_url,
-        aws_access_key_id=settings.storage_access_key,
-        aws_secret_access_key=settings.storage_secret_key,
-    )
-
-
-def ensure_bucket_exists() -> None:
-    client = get_s3_client()
-    try:
-        client.head_bucket(Bucket=settings.storage_bucket)
-    except ClientError as e:
-        if e.response["Error"]["Code"] in ("404", "NoSuchBucket"):
-            client.create_bucket(Bucket=settings.storage_bucket)
-        else:
-            raise
-
-
-def pcm_to_wav(pcm_data: bytes, sample_rate: int = 24000) -> bytes:
-    """Wrap raw 16-bit mono PCM data in a WAV container."""
-    buf = io.BytesIO()
-    with wave.open(buf, "wb") as wf:
-        wf.setnchannels(1)
-        wf.setsampwidth(2)  # 16-bit
-        wf.setframerate(sample_rate)
-        wf.writeframes(pcm_data)
-    return buf.getvalue()
-
-
-def upload_audio(object_key: str, audio_bytes: bytes, content_type: str = "audio/wav") -> None:
-    client = get_s3_client()
-    client.put_object(
-        Bucket=settings.storage_bucket,
-        Key=object_key,
-        Body=audio_bytes,
-        ContentType=content_type,
-    )
-
-
-def download_audio(object_key: str) -> tuple[bytes, str]:
-    """Return (file_bytes, content_type)."""
-    client = get_s3_client()
-    response = client.get_object(Bucket=settings.storage_bucket, Key=object_key)
-    content_type = response.get("ContentType", "audio/wav")
-    return response["Body"].read(), content_type

api/docs/design-doc-object-storage.md

@@ -0,0 +1,82 @@
+# Design Document: Object Storage with Bunny CDN
+
+This is a technical design document for implementing object (e.g. audio file) storage with Bunny CDN.  This directory (`api/docs`) contains other similar files, notably `architecture.md` and `domain.md`.  When you have worked through the change described here, please update `architecture.md`
+
+## The problem
+
+Language Learning App has audio as a core component, which requires files to be delivered to the end user.  When developing locally, these files have been stored in a min.io service, mimicking an S3-like storage bucket.  
+
+Using this approach on a deployed instance (e.g. on a VPS using Docker), would result in high bandwidth and therefore a high cost.  Using a dedicated, EU-based service like Bunny allows us to offload the delivery of content to a third-party at reduced cost (great!)
+
+## The current implementation
+
+Object storage was one of the first features built into this software in MVP state, as such it does not fit within the current architecture.
+
+Right now `api/app/storage.py` contains some helper functions, notably the `upload_audio` and `download_audio` functions.
+
+Users (through the web client) retrieve the media through two URLs (detailed in `api/app/routers/media.py`):
+
+- `GET /media/adventure-audio/{filename:path}` for the choose-your-own-adventure file names
+- `GET /media/{filename:path}`, used for the summary transcriptions
+
+## The solution
+
+We are going to use Bunny (bunny.net) as the CDN for all objects in deployed environments (right now, just production — in the future preprod or staging may exist).
+
+Locally, for development purposes, we retain the use of MinIO. To decide which backend to use, we introduce an environment variable `STORAGE_PROVIDER` with a default value of `local` and an accepted alternative of `bunny`.
+
+In situations where we use `local`, the existing `/media/..` proxy endpoints are returned when constructing audio URLs (e.g. in `api/app/routers/bff/articles.py` and `api/app/routers/bff/adventure.py`). When we use `bunny`, the Bunny CDN URL is returned directly so the request is never proxied through our service.
+
+### Client interface
+
+We will create a `BunnyClient` in `api/app/outbound/bunny/bunny_client.py` and extract the current MinIO logic into a `MinioClient` in `api/app/outbound/minio/minio_client.py`. Both implement a shared `StorageClient` protocol.
+
+The interface is **generic** — the clients are storage adapters and must not encode domain concepts. Path construction (which directory, which filename) is the responsibility of the caller (the service layer), not the client.
+
+```python
+class StorageClient(Protocol):
+    def upload(self, path: str, data: bytes) -> bool: ...
+    def get_url(self, path: str) -> str: ...
+    def delete(self, path: str) -> bool: ...
+```
+
+Services construct paths using hardcoded directory prefixes (e.g. `"adventure-audio/"`, `"audio/"`). These are constants, not environment variables — they are not environment-specific and do not belong in config.
+
+### Factory and instantiation
+
+A factory function reads `STORAGE_PROVIDER` and returns the appropriate `StorageClient` implementation. The client is instantiated **once at app startup** (e.g. in `main.py`) as a module-level singleton — not per-request. This is consistent with how other outbound clients (`AnthropicClient`, `GeminiClient`, etc.) are handled.
+
+### Bunny configuration
+
+Bunny requires the following environment variables:
+
+- `BUNNY_ZONE` — the storage zone name (the zone `languagelearningapp` has been created in the Bunny UI). No "DEFAULT" suffix; there is one zone.
+- `BUNNY_API_KEY` — the Bunny API key for upload/delete operations.
+- `BUNNY_CDN_BASE_URL` — the public CDN hostname used to construct delivery URLs.
+
+### Signed vs. public URLs
+
+Audio files are user-specific (i.e. one user should not be able to use another user's audio URL), Bunny signed URLs are required. Public CDN URLs are shareable by anyone who has the link.
+
+As per Bunny's own documentation they recommend the token.py package:
+
+```py
+from token import sign_url
+
+url = sign_url(
+    "https://myzone.b-cdn.net/videos/stream1/playlist.m3u8",
+    "your-security-key",
+    expiration_time=3600,
+    is_directory=True,
+    path_allowed="/videos/stream1/",
+    countries_allowed="GB",
+)
+```
+
+`get_url(path)` on the `BunnyClient` must generate a time-limited (pick a sensible default for audio content here) signed URL using the Bunny Token Authentication feature. The MinIO implementation would use pre-signed S3 URLs for consistency.
+
+Create a sibling method that explicitely creates public URLs for any future public content, call this `get_public_url`.
+
+### Misc
+
+`pcm_to_wav()` currently lives in `api/app/storage.py` but is a Gemini output concern. Move it to the Gemini client module (`api/app/outbound/gemini/`) when carrying out this refactor.

api/docs/technical-doc-choose-your-own-adventure.md

@@ -146,7 +146,6 @@ def parse_llm_response(text: str) -> tuple[str, list[tuple[str, str]], str]:
 app/domain/models/adventure.py
 app/domain/services/adventure_service.py
 app/routers/api/adventures.py
-app/routers/bff/adventures.py
 app/outbound/postgres/entities/adventure_entities.py
 app/outbound/postgres/repositories/adventure_repository.py
 alembic/versions/20260503_0016_add_choose_your_own_adventure.py
@@ -158,7 +157,6 @@ Modified files:
 ```
 app/outbound/anthropic/anthropic_client.py  (add 2 methods)
 app/routers/api/main.py                     (register router)
-app/routers/bff/main.py                     (register router)
 ```
 
 ---

docker-compose-dev.yml

@@ -0,0 +1,80 @@
+services:
+  db:
+    image: postgres:16-alpine
+    environment:
+      POSTGRES_USER: ${POSTGRES_USER:-langlearn}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
+      POSTGRES_DB: ${POSTGRES_DB:-langlearn}
+    volumes:
+      - pgdata:/var/lib/postgresql/data
+    ports:
+      - "5432:5432"
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER:-langlearn}"]
+      interval: 5s
+      timeout: 5s
+      retries: 10
+
+  storage:
+    image: minio/minio:latest
+    command: server /data --console-address ":9001"
+    environment:
+      MINIO_ROOT_USER: ${STORAGE_ACCESS_KEY:-langlearn}
+      MINIO_ROOT_PASSWORD: ${STORAGE_SECRET_KEY}
+    ports:
+      - "9000:9000"
+      - "9001:9001"
+    volumes:
+      - storagedata:/data
+    healthcheck:
+      test: ["CMD-SHELL", "curl -sf http://localhost:9000/minio/health/live || exit 1"]
+      interval: 5s
+      timeout: 5s
+      retries: 10
+
+  api:
+    build: ./api
+    volumes: 
+      - ./api:/app:z
+    ports:
+      - "${API_PORT:-8000}:8000"
+    command: uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
+    environment:
+      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-langlearn}:${POSTGRES_PASSWORD}@db:5432/${POSTGRES_DB:-langlearn}
+      ADMIN_USER_EMAILS: ${ADMIN_USER_EMAILS:-wilson@thomaswilson.xyz}
+      API_BASE_URL: ${API_BASE_URL:-http://localhost:8000}
+      JWT_SECRET: ${JWT_SECRET}
+      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY}
+      DEEPL_API_KEY: ${DEEPL_API_KEY}
+      DEEPGRAM_API_KEY: ${DEEPGRAM_API_KEY}
+      GEMINI_API_KEY: ${GEMINI_API_KEY}
+      PYTHONPATH: /app
+      STORAGE_PROVIDER: local
+      STORAGE_ENDPOINT_URL: http://storage:9000
+      STORAGE_ACCESS_KEY: ${STORAGE_ACCESS_KEY:-langlearn}
+      STORAGE_SECRET_KEY: ${STORAGE_SECRET_KEY}
+      STORAGE_BUCKET: ${STORAGE_BUCKET:-langlearn}
+      TRANSACTIONAL_EMAIL_PROVIDER: ${TRANSACTIONAL_EMAIL_PROVIDER:-stub}
+    depends_on:
+      db:
+        condition: service_healthy
+      storage:
+        condition: service_healthy
+    restart: unless-stopped
+
+  frontend:
+    build:
+      context: ./frontend
+      args:
+        PUBLIC_API_BASE_URL: ${PUBLIC_API_BASE_URL:-http://localhost:8000}
+    ports:
+      - "${FRONTEND_PORT:-3000}:3000"
+    environment:
+      ORIGIN: ${ORIGIN:-http://localhost:3000}
+    depends_on:
+      - api
+    restart: unless-stopped
+
+volumes:
+  pgdata:
+  storagedata:

docker-compose-prod.yml

@@ -0,0 +1,83 @@
+services:
+  db:
+    image: postgres:16-alpine
+    environment:
+      POSTGRES_USER: ${POSTGRES_USER:-langlearn}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
+      POSTGRES_DB: ${POSTGRES_DB:-langlearn}
+    volumes:
+      - pgdata:/var/lib/postgresql/data
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER:-langlearn}"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+    restart: unless-stopped
+    deploy:
+      resources:
+        limits:
+          cpus: '1'
+          memory: 1G
+
+  api:
+    build: ./api
+    ports:
+      - "${API_PORT:-8000}:8000"
+    command: uvicorn app.main:app --host 0.0.0.0 --port 8000 --workers 2
+    environment:
+      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-langlearn}:${POSTGRES_PASSWORD}@db:5432/${POSTGRES_DB:-langlearn}
+      ADMIN_USER_EMAILS: ${ADMIN_USER_EMAILS}
+      API_BASE_URL: ${API_BASE_URL}
+      JWT_SECRET: ${JWT_SECRET}
+      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY}
+      DEEPL_API_KEY: ${DEEPL_API_KEY}
+      DEEPGRAM_API_KEY: ${DEEPGRAM_API_KEY}
+      GEMINI_API_KEY: ${GEMINI_API_KEY}
+      PYTHONPATH: /app
+      STORAGE_PROVIDER: bunny
+      BUNNY_ZONE: ${BUNNY_ZONE}
+      BUNNY_API_KEY: ${BUNNY_API_KEY}
+      BUNNY_CDN_BASE_URL: ${BUNNY_CDN_BASE_URL}
+      BUNNY_TOKEN_AUTH_KEY: ${BUNNY_TOKEN_AUTH_KEY}
+      TRANSACTIONAL_EMAIL_PROVIDER: ${TRANSACTIONAL_EMAIL_PROVIDER}
+    healthcheck:
+      test: ["CMD-SHELL", "curl -sf http://localhost:8000/health || exit 1"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+      start_period: 20s
+    depends_on:
+      db:
+        condition: service_healthy
+    restart: unless-stopped
+    deploy:
+      resources:
+        limits:
+          cpus: '1'
+          memory: 1G
+
+  frontend:
+    build:
+      context: ./frontend
+      args:
+        PUBLIC_API_BASE_URL: ${PUBLIC_API_BASE_URL}
+    ports:
+      - "${FRONTEND_PORT:-3000}:3000"
+    environment:
+      ORIGIN: ${ORIGIN}
+    depends_on:
+      api:
+        condition: service_healthy
+    restart: unless-stopped
+    deploy:
+      resources:
+        limits:
+          cpus: '0.5'
+          memory: 256M
+
+volumes:
+  pgdata:
+
+networks:
+  default:
+    name: langlearn

frontend/src/routes/app/adventures/[id]/+page.svelte

@@ -299,6 +299,7 @@
 		<LatestEntry
 			sourceText={latestEntry?.story_text}
 			translationText={latestEntry?.translation}
+			storyTextLinguisticData={latestEntry?.story_text_linguistic_data}
 			audioUrl={latestEntry?.audio_url}
 			onSelectNextStep={handleNextStepSelect}
 			isWaitingForGeneration={$adventureState.ui.isWaitingForGeneration}

frontend/src/routes/app/adventures/[id]/LatestEntry.svelte

@@ -6,6 +6,7 @@
 	type Props = {
 		sourceText: string | null | undefined;
 		translationText: string | null | undefined;
+		storyTextLinguisticData: Record<string, unknown> | null | undefined;
 		audioUrl: string | null | undefined;
 
 		onSelectNextStep: (optionId: string) => Promise<void>;
@@ -24,6 +25,7 @@
 	const {
 		sourceText,
 		translationText,
+		storyTextLinguisticData,
 		audioUrl,
 
 		onSelectNextStep,
@@ -33,10 +35,222 @@
 		errorMessage
 	}: Props = $props();
 
-	const sourceParagraphs = $derived.by(() => toParagraphs(sourceText));
+	type LinguisticToken = {
-	const translationParagraphs = $derived.by(() => toParagraphs(translationText));
+		text: string;
+		lemma: string | null;
+		pos: string | null;
+	};
 
-	let lastClickedParagraphIndex: number | null = $state(null);
+	type TextSegment = {
+		kind: 'text';
+		text: string;
+	};
+
+	type WordSegment = {
+		kind: 'word';
+		text: string;
+		lemma: string | null;
+		pos: string | null;
+	};
+
+	type SentenceSegments = {
+		key: string;
+		sourceText: string;
+		targetText: string;
+		sourceSegments: Array<TextSegment | WordSegment>;
+		targetSegments: Array<TextSegment | WordSegment>;
+	};
+
+	type LinguisticParagraph = {
+		key: string;
+		sourceText: string;
+		targetText: string;
+		sentences: SentenceSegments[];
+	};
+
+	function isRecord(value: unknown): value is Record<string, unknown> {
+		return typeof value === 'object' && value !== null;
+	}
+
+	function asString(value: unknown): string | null {
+		return typeof value === 'string' ? value : null;
+	}
+
+	function parseTokens(value: unknown): LinguisticToken[] {
+		if (!Array.isArray(value)) {
+			return [];
+		}
+
+		return value
+			.map((token): LinguisticToken | null => {
+				if (!isRecord(token)) {
+					return null;
+				}
+
+				const text = asString(token.text);
+				if (!text) {
+					return null;
+				}
+
+				return {
+					text,
+					lemma: asString(token.lemma),
+					pos: asString(token.pos)
+				};
+			})
+			.filter((token): token is LinguisticToken => token !== null);
+	}
+
+	function buildSegments(
+		text: string,
+		tokens: LinguisticToken[]
+	): Array<TextSegment | WordSegment> {
+		if (tokens.length === 0) {
+			return text ? [{ kind: 'text', text }] : [];
+		}
+
+		if (!text) {
+			return tokens.flatMap((token, index) => [
+				{ kind: 'word', text: token.text, lemma: token.lemma, pos: token.pos } as WordSegment,
+				...(index < tokens.length - 1 ? ([{ kind: 'text', text: ' ' }] as TextSegment[]) : [])
+			]);
+		}
+
+		const segments: Array<TextSegment | WordSegment> = [];
+		let cursor = 0;
+
+		for (const token of tokens) {
+			const tokenIndex = text.indexOf(token.text, cursor);
+
+			if (tokenIndex === -1) {
+				continue;
+			}
+
+			if (tokenIndex > cursor) {
+				segments.push({
+					kind: 'text',
+					text: text.slice(cursor, tokenIndex)
+				});
+			}
+
+			segments.push({
+				kind: 'word',
+				text: token.text,
+				lemma: token.lemma,
+				pos: token.pos
+			});
+
+			cursor = tokenIndex + token.text.length;
+		}
+
+		if (cursor < text.length) {
+			segments.push({
+				kind: 'text',
+				text: text.slice(cursor)
+			});
+		}
+
+		return segments.length > 0 ? segments : [{ kind: 'text', text }];
+	}
+
+	function parseLinguisticParagraphs(
+		value: Record<string, unknown> | null | undefined
+	): LinguisticParagraph[] {
+		if (!value) {
+			return [];
+		}
+
+		const paragraphs = value.paragraphs;
+		if (!Array.isArray(paragraphs)) {
+			return [];
+		}
+
+		return paragraphs
+			.map((paragraphValue, paragraphIndex): LinguisticParagraph | null => {
+				if (!isRecord(paragraphValue)) {
+					return null;
+				}
+
+				const sentencesRaw = paragraphValue.sentences;
+				const sentenceValues = Array.isArray(sentencesRaw) ? sentencesRaw : [];
+
+				const sentences = sentenceValues
+					.map((sentenceValue, sentenceIndex): SentenceSegments | null => {
+						if (!isRecord(sentenceValue)) {
+							return null;
+						}
+
+						const sourceSentence = asString(sentenceValue.source_text) ?? '';
+						const targetSentence = asString(sentenceValue.target_text) ?? '';
+						const sourceTokens = parseTokens(sentenceValue.source_tokens);
+						const targetTokens = parseTokens(sentenceValue.target_tokens);
+
+						return {
+							key: `${paragraphIndex}-${sentenceIndex}`,
+							sourceText: sourceSentence,
+							targetText: targetSentence,
+							sourceSegments: buildSegments(sourceSentence, sourceTokens),
+							targetSegments: buildSegments(targetSentence, targetTokens)
+						};
+					})
+					.filter((sentence): sentence is SentenceSegments => sentence !== null);
+
+				const sourceText =
+					asString(paragraphValue.source_text) ??
+					sentences
+						.map((sentence) => sentence.sourceText)
+						.filter(Boolean)
+						.join(' ');
+				const targetText =
+					asString(paragraphValue.target_text) ??
+					sentences
+						.map((sentence) => sentence.targetText)
+						.filter(Boolean)
+						.join(' ');
+
+				if (sentences.length === 0 && (sourceText || targetText)) {
+					sentences.push({
+						key: `${paragraphIndex}-0`,
+						sourceText,
+						targetText,
+						sourceSegments: sourceText ? [{ kind: 'text', text: sourceText }] : [],
+						targetSegments: targetText ? [{ kind: 'text', text: targetText }] : []
+					});
+				}
+
+				if (!sourceText && !targetText && sentences.length === 0) {
+					return null;
+				}
+
+				return {
+					key: `p-${paragraphIndex}`,
+					sourceText,
+					targetText,
+					sentences
+				};
+			})
+			.filter((paragraph): paragraph is LinguisticParagraph => paragraph !== null);
+	}
+
+	function isWordLike(text: string): boolean {
+		return /[\p{L}\p{N}]/u.test(text);
+	}
+
+	const linguisticParagraphs = $derived.by(() =>
+		parseLinguisticParagraphs(storyTextLinguisticData)
+	);
+	const sourceParagraphs = $derived.by(() =>
+		linguisticParagraphs.length > 0
+			? linguisticParagraphs.map((paragraph) => paragraph.targetText).filter(Boolean)
+			: toParagraphs(sourceText)
+	);
+	const translationParagraphs = $derived.by(() =>
+		linguisticParagraphs.length > 0
+			? linguisticParagraphs.map((paragraph) => paragraph.sourceText).filter(Boolean)
+			: toParagraphs(translationText)
+	);
+
+	let selectedWord: { sentenceKey: string; text: string } | null = $state(null);
 	let sourcePane = $state<HTMLDivElement | undefined>();
 	let translationPane = $state<HTMLDivElement | undefined>();
 	let suppressSourceScroll = $state(false);
@@ -101,8 +315,9 @@
 		}, 20000);
 	}
 
-	function handleParagraphClicked(paragraphIndex: number) {
+	function handleWordClicked(sentenceKey: string, text: string) {
-		lastClickedParagraphIndex = paragraphIndex;
+		selectedWord = { sentenceKey, text };
+		showTranslation();
 	}
 
 	async function handleNextStepSelect(optionId: string) {
@@ -146,18 +361,45 @@
 			<div class="pane source-pane">
 				<div class="latest-entry__pane-body" bind:this={sourcePane} onscroll={handleSourceScroll}>
 					{#if sourceParagraphs.length > 0}
-						{#each sourceParagraphs as paragraph, index (index)}
+						{#if linguisticParagraphs.length > 0}
-							<button
+							{#each linguisticParagraphs as paragraph (paragraph.key)}
-								type="button"
+								<p class="paragraph paragraph--text" data-language="source">
-								class="paragraph"
+									{#each paragraph.sentences as sentence (sentence.key)}
-								class:active={lastClickedParagraphIndex === index}
+										<span
-								data-paragraph-index={index}
+											class="sentence-chunk"
-								data-language="source"
+											class:active-sentence={selectedWord?.sentenceKey === sentence.key}
-								onclick={() => handleParagraphClicked(index)}
+										>
-							>
+											{#each sentence.targetSegments as segment, segmentIndex (`${sentence.key}-target-${segmentIndex}`)}
-								{paragraph}
+												{#if segment.kind === 'word' && isWordLike(segment.text)}
-							</button>
+													<button
-						{/each}
+														type="button"
+														class="word-token"
+														class:active={selectedWord?.sentenceKey === sentence.key &&
+															selectedWord?.text === segment.text}
+														title={segment.lemma ? `Lemma: ${segment.lemma}` : undefined}
+														onclick={() => handleWordClicked(sentence.key, segment.text)}
+													>
+														{segment.text}
+													</button>
+												{:else}
+													<span>{segment.text}</span>
+												{/if}
+											{/each}
+										</span>
+									{/each}
+								</p>
+							{/each}
+						{:else}
+							{#each sourceParagraphs as paragraph, index (index)}
+								<p
+									class="paragraph paragraph--text"
+									data-paragraph-index={index}
+									data-language="source"
+								>
+									{paragraph}
+								</p>
+							{/each}
+						{/if}
 					{:else}
 						<div class="loading-block" role="status" aria-live="polite">
 							<p class="loading-block__label">{statusMessage || 'Writing your next entry...'}</p>
@@ -177,6 +419,12 @@
 					</button>
 				</header>
 
+				{#if selectedWord}
+					<p class="translation-selected-word" role="status" aria-live="polite">
+						Selected word: <strong>{selectedWord.text}</strong>
+					</p>
+				{/if}
+
 				{#if translationVisible}
 					<div
 						class="latest-entry__pane-body"
@@ -184,18 +432,42 @@
 						onscroll={handleTranslationScroll}
 					>
 						{#if translationParagraphs.length > 0}
-							{#each translationParagraphs as paragraph, index (index)}
+							{#if linguisticParagraphs.length > 0}
-								<button
+								{#each linguisticParagraphs as paragraph (paragraph.key)}
-									type="button"
+									<p class="paragraph paragraph--text" data-language="translation">
-									class="paragraph"
+										{#each paragraph.sentences as sentence (sentence.key)}
-									class:active={lastClickedParagraphIndex === index}
+											<span
-									data-paragraph-index={index}
+												class="sentence-chunk"
-									data-language="translation"
+												class:active-sentence={selectedWord?.sentenceKey === sentence.key}
-									onclick={() => handleParagraphClicked(index)}
+											>
-								>
+												{#each sentence.sourceSegments as segment, segmentIndex (`${sentence.key}-source-${segmentIndex}`)}
-									{paragraph}
+													{#if segment.kind === 'word'}
-								</button>
+														<span
-							{/each}
+															class="word-token word-token--passive"
+															class:active={selectedWord?.sentenceKey === sentence.key &&
+																selectedWord?.text === segment.text}
+														>
+															{segment.text}
+														</span>
+													{:else}
+														<span>{segment.text}</span>
+													{/if}
+												{/each}
+											</span>
+										{/each}
+									</p>
+								{/each}
+							{:else}
+								{#each translationParagraphs as paragraph, index (index)}
+									<p
+										class="paragraph paragraph--text"
+										data-paragraph-index={index}
+										data-language="translation"
+									>
+										{paragraph}
+									</p>
+								{/each}
+							{/if}
 						{:else}
 							<div class="loading-block" role="status" aria-live="polite">
 								<p class="loading-block__label">
@@ -511,6 +783,18 @@
 		color: color-mix(in srgb, var(--color-on-surface) 72%, transparent);
 	}
 
+	.translation-selected-word {
+		margin: 0;
+		padding: 0 var(--latest-entry-pane-padding) var(--space-2);
+		font-size: var(--text-label-md);
+		color: color-mix(in srgb, var(--color-on-surface) 84%, transparent);
+	}
+
+	.translation-selected-word strong {
+		font-weight: var(--weight-semibold);
+		color: var(--color-primary);
+	}
+
 	.latest-entry__pane-body::-webkit-scrollbar {
 		width: 0.75rem;
 	}
@@ -542,9 +826,58 @@
 		transition: background-color var(--duration-fast) var(--ease-standard);
 	}
 
-	.paragraph.active {
+	.paragraph--text {
-		background-color: color-mix(in srgb, var(--color-primary-container) 56%, transparent);
+		margin: 0;
-		border-radius: var(--radius-md);
+		padding: 0;
+		border: none;
+		background: transparent;
+		cursor: default;
+		white-space: pre-wrap;
+	}
+
+	.sentence-chunk {
+		display: inline;
+		border-radius: var(--radius-sm);
+		transition: background-color var(--duration-fast) var(--ease-standard);
+
+		&::after {
+			content: '';
+			display: inline-block;
+			width: 0.85ch;
+		}
+	}
+
+	.sentence-chunk.active-sentence {
+		background-color: color-mix(in srgb, var(--color-primary-container) 32%, transparent);
+	}
+
+	.word-token {
+		display: inline;
+		padding: 0;
+		margin: 0;
+		border: none;
+		background: transparent;
+		color: inherit;
+		font: inherit;
+		line-height: inherit;
+		cursor: pointer;
+	}
+
+	.word-token:hover {
+		text-decoration: underline;
+	}
+
+	.word-token.active {
+		color: var(--color-primary);
+		text-decoration: underline;
+	}
+
+	.word-token--passive {
+		cursor: default;
+	}
+
+	.word-token--passive:hover {
+		text-decoration: none;
 	}
 
 	.paragraph:focus-visible {

frontend/src/routes/app/adventures/[id]/adventureState.ts

@@ -4,6 +4,7 @@ export type AdventureEntry = {
 	id: string;
 	story_text: string | null;
 	translation: string | null;
+	story_text_linguistic_data: Record<string, unknown> | null;
 	audio_url: string | null;
 	generated_from_choice_id: string | null;
 	possible_choices: { id: string; text: string }[] | null;