Phase 4: retire Streamlit, Docker web app, answer quality, final README

.dockerignore

@@ -10,6 +10,4 @@ chroma_db
 legacy
 tests
 docs
-components
-styles
-utils
+web

DEPLOY.md

@@ -62,7 +62,7 @@ You create every account and paste every secret yourself. Nothing secret goes in
    | Branch | `main` |
    | Region | **Frankfurt (EU Central)** |
    | Instance type | **Free** |
-   | Dockerfile Path (Advanced) | `Dockerfile.api` (the default `./Dockerfile` builds the Streamlit app instead) |
+   | Dockerfile Path (Advanced) | `Dockerfile.api` |
    | Docker Build Context | `.` |
    | Health Check Path (Advanced) | `/ping` |
    | Auto-Deploy | On Commit |
@@ -140,7 +140,7 @@ without it, all visitors would share one rate limit.
    |---|---|
    | Repository | this repository |
    | Branch | `main` |
-   | Main file path | `app.py` |
+   | Main file path | `legacy/streamlit/app.py` |
    | App URL | a short name, e.g. `floatchat` |
 
 4. Open **Advanced settings**: choose Python **3.11** or newer, and paste into **Secrets**:
@@ -150,7 +150,7 @@ without it, all visitors would share one rate limit.
    FRONTEND_TOKEN = "<the same secret as on Render>"
    ```
 
-5. Click **Deploy**. Streamlit installs `requirements.txt` (frontend packages only) and starts the app in a few
+5. Click **Deploy**. Streamlit installs `legacy/streamlit/requirements.txt` (frontend packages only) and starts the app in a few
    minutes. The sidebar should show green dots for PostgreSQL, the vector index and the LLM.
 
 ## Troubleshooting

docker-compose.yml

@@ -1,5 +1,5 @@
 # One-command local stack: docker compose up --build
-# App: http://localhost:8501   API docs: http://localhost:8000/docs
+# Web app: http://localhost:3000   API docs: http://localhost:8000/docs
 services:
   db:
     image: pgvector/pgvector:pg16
@@ -32,7 +32,7 @@ services:
       ADMIN_DATABASE_URL: postgresql+psycopg://floatchat:floatchat@db:5432/argo
       DATABASE_URL: postgresql+psycopg://floatchat_reader:local-reader-password@db:5432/argo
       READER_PASSWORD: local-reader-password
-      ALLOWED_ORIGINS: http://localhost:8501,http://localhost:3000
+      ALLOWED_ORIGINS: http://localhost:3000
     ports:
       - "8000:8000"
     extra_hosts:
@@ -41,12 +41,15 @@ services:
       db:
         condition: service_healthy
 
-  frontend:
-    build: .
+  web:
+    build:
+      context: ./web
+      args:
+        NEXT_PUBLIC_API_URL: http://localhost:8000   # what your browser uses
     environment:
-      BACKEND_URL: http://backend:8000
+      API_URL_INTERNAL: http://backend:8000          # what the web server uses inside Docker
     ports:
-      - "8501:8501"
+      - "3000:3000"
     depends_on:
       - backend
 

floatchat/answer.py

@@ -135,7 +135,8 @@ def _analytical_answer(route: Route, agent: SqlAgent, writer: _Writer, step) ->
     writer.write(f"**{title}**\n\n{format_results_for_display(df, intent)}")
     insight = writer.llm(
         f"Question: {route.question}\n\nSQL result (first rows):\n{df.head(15).to_string(index=False)}\n\n"
-        "In 2-3 sentences, explain the main oceanographic pattern in this result.",
+        "In 2-3 sentences of plain prose, explain the main oceanographic pattern in this result. Do not use a list, "
+        "table or heading, and do not restate the rows one by one.",
         prefix="\n\n**What this means:** ",
     )
     return {

floatchat/config.py

@@ -36,7 +36,7 @@ def _sqlalchemy_url(value: str) -> str:
 FRONTEND_TOKEN = os.getenv("FRONTEND_TOKEN", "")
 ALLOWED_ORIGINS = [
     origin.strip()
-    for origin in os.getenv("ALLOWED_ORIGINS", "http://localhost:3000,http://localhost:8501").split(",")
+    for origin in os.getenv("ALLOWED_ORIGINS", "http://localhost:3000").split(",")
     if origin.strip()
 ]
 

floatchat/llm.py

@@ -27,6 +27,7 @@ def is_configured() -> bool:
     return True
 
 
+# Reasoning models (such as gpt-oss) spend part of max_tokens thinking before they answer, so the limits leave room
 def _payload(prompt: str, system: str | None, temperature: float, max_tokens: int, json_mode: bool,
              stream: bool = False) -> dict:
     messages = [{"role": "system", "content": system}] if system else []
@@ -58,7 +59,7 @@ def _post(payload: dict, stream: bool = False):
         _sleep(min(wait, MAX_RETRY_WAIT))
 
 
-def chat(prompt: str, system: str | None = None, temperature: float = 0.2, max_tokens: int = 600,
+def chat(prompt: str, system: str | None = None, temperature: float = 0.2, max_tokens: int = 1200,
          json_mode: bool = False) -> str | None:
     if not is_configured():
         return None
@@ -71,7 +72,7 @@ def chat(prompt: str, system: str | None = None, temperature: float = 0.2, max_t
 
 
 def chat_stream(prompt: str, system: str | None = None, temperature: float = 0.2,
-                max_tokens: int = 600) -> Iterator[str]:
+                max_tokens: int = 1200) -> Iterator[str]:
     """Yield the answer text as the LLM writes it; yields nothing if no LLM is available or the call fails."""
     if not is_configured():
         return

floatchat/router.py

@@ -74,7 +74,7 @@ def _prompt(question: str, history: list[dict]) -> str:
 
 
 def llm_route(question: str, history: list[dict], chat) -> Route | None:
-    reply = chat(_prompt(question, history or []), temperature=0.0, max_tokens=300, json_mode=True)
+    reply = chat(_prompt(question, history or []), temperature=0.0, max_tokens=600, json_mode=True)
     if not reply:
         return None
     try:

legacy/streamlit/README.md

@@ -0,0 +1,15 @@
+# Retired Streamlit apps
+
+FloatChat's frontend is now the Next.js app in [`web/`](../../web). These are kept for reference:
+
+- `app.py`: the interim Streamlit app that served as the showcase before the web app (Phases 1 and 2). It still works
+  against the API:
+
+  ```bash
+  pip install -r legacy/streamlit/requirements.txt
+  BACKEND_URL=http://localhost:8000 streamlit run legacy/streamlit/app.py
+  ```
+
+  On a hosted deployment it passes each visitor's IP with a shared `FRONTEND_TOKEN`, which the API still accepts.
+- `streamlit_app.py`, `components/`, `styles/`, `utils/`: an earlier experimental dashboard
+  (see `README_DASHBOARD.md`), with its tests in `tests/dashboard/`.

pyproject.toml

@@ -1,7 +1,7 @@
 [tool.ruff]
 line-length = 120
 target-version = "py311"
-extend-exclude = ["legacy", "components", "styles", "utils", "tests/dashboard", "streamlit_app.py", "dashboard_config.py", "app.py"]
+extend-exclude = ["legacy", "web"]
 
 [tool.ruff.lint]
 select = ["E", "F", "I", "B"]