Add LLM_PROVIDER preset (groq/gemini/openrouter) for the free LLM

Lets the answer provider be chosen with one variable instead of typing the
base URL and model. Gemini and OpenRouter offer Google/email sign-in for
users who can't sign up via GitHub. LLM_BASE_URL / LLM_MODEL still override
the preset. Docs updated.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WN4QRr6dTE2W7hQ2SDnLmY

README.md

@@ -40,12 +40,12 @@ all runnable with one command and deployable as a single container.
   the retrieval is transparent.
 - **Hybrid retrieval** — BM25 (`rank_bm25`) combined with a FAISS cosine search
   over TF-IDF vectors; deliberately lightweight, no heavyweight model.
-- **Optional LLM answers (free)** — set `LLM_API_KEY` (defaults to
-  [Groq](https://console.groq.com/keys), a free OpenAI-compatible API) and
-  answers become grounded summaries written from the retrieved passages (the
-  source passage is still shown). Point `LLM_BASE_URL`/`LLM_MODEL` at any
-  OpenAI-compatible provider. Without a key it falls back to an extractive
-  answer, so the app runs with no credentials and no cost.
+- **Optional LLM answers (free)** — set `LLM_PROVIDER` (`groq` | `gemini` |
+  `openrouter`) and `LLM_API_KEY`, and answers become grounded summaries written
+  from the retrieved passages (the source passage is still shown). All three are
+  free OpenAI-compatible providers with email/Google sign-in. Without a key it
+  falls back to an extractive answer, so the app runs with no credentials and no
+  cost.
 - **Feedback & telemetry** — each question is logged to SQLite with its measured
   latency; 👍/👎 feedback and a `/dashboard` page show total questions, median
   latency, and thumbs-up rate live from the DB.
@@ -134,10 +134,12 @@ The API is a standard uvicorn app. On **Railway**:
    [`backend/Dockerfile`](./backend/Dockerfile) (which installs faiss's
    `libgomp1` and honors Railway's injected `$PORT`).
 3. **Settings → Networking → Generate Domain** to get a public URL.
-4. *(Optional)* In the service **Variables**, add `LLM_API_KEY` (a free
-   [Groq](https://console.groq.com/keys) key) to enable LLM-written answers
-   (otherwise the API serves extractive answers). Add a **Volume** at `/data` +
-   `DOCUASK_DB=/data/docuask.db` to persist telemetry across restarts.
+4. *(Optional)* In the service **Variables**, set `LLM_PROVIDER` (e.g. `gemini`)
+   and `LLM_API_KEY` (a free key — Gemini's is at
+   [aistudio.google.com/apikey](https://aistudio.google.com/apikey), Google
+   sign-in) to enable LLM-written answers; otherwise the API serves extractive
+   answers. Add a **Volume** at `/data` + `DOCUASK_DB=/data/docuask.db` to
+   persist telemetry across restarts.
 5. Copy the public URL — you'll set it as `DOCUASK_API_URL` in the HF Space
    (step 3 above).
 

backend/.env.example

@@ -16,16 +16,14 @@
 # DocuAsk falls back to the extractive best-sentence answer (no key, no cost,
 # still shows the source passage).
 #
-# Default provider is Groq (free, no credit card): get a key at
-# https://console.groq.com/keys
-# LLM_API_KEY=gsk_...
+# Pick a free provider with LLM_PROVIDER (groq | gemini | openrouter) and set
+# LLM_API_KEY. All three offer email/Google sign-in — no GitHub needed.
+#   groq       – https://console.groq.com/keys        (default)
+#   gemini     – https://aistudio.google.com/apikey   (sign in with Google)
+#   openrouter – https://openrouter.ai/keys
+# LLM_PROVIDER=gemini
+# LLM_API_KEY=...
 #
-# Override to use a different free provider:
-#   Groq (default):  LLM_BASE_URL=https://api.groq.com/openai/v1
-#                    LLM_MODEL=llama-3.3-70b-versatile
-#   Google Gemini:   LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
-#                    LLM_MODEL=gemini-2.0-flash
-#   OpenRouter:      LLM_BASE_URL=https://openrouter.ai/api/v1
-#                    LLM_MODEL=<a free model id>
+# Advanced: override the preset's endpoint/model directly.
 # LLM_BASE_URL=https://api.groq.com/openai/v1
 # LLM_MODEL=llama-3.3-70b-versatile

backend/app/generate.py

@@ -8,8 +8,11 @@ the app runs with no network access and no credentials.
 
 Configure via env vars:
   LLM_API_KEY   – your free API key (required to enable generation)
-  LLM_BASE_URL  – OpenAI-compatible base URL (default: Groq)
-  LLM_MODEL     – model id (default: a current Groq Llama model)
+  LLM_PROVIDER  – preset base URL + model: "groq" (default), "gemini",
+                  or "openrouter". Sign-in for each is by email/Google, no
+                  GitHub required.
+  LLM_BASE_URL  – override the preset's OpenAI-compatible base URL
+  LLM_MODEL     – override the preset's model id
 
 The OpenAI SDK is imported lazily so the package is never a hard import-time
 dependency of the API.
@@ -22,10 +25,24 @@ import os
 
 logger = logging.getLogger("docuask")
 
-# Defaults target Groq's free, OpenAI-compatible API. Override any of these to
-# use a different free provider (e.g. Google Gemini's OpenAI endpoint).
-_BASE_URL = os.getenv("LLM_BASE_URL", "https://api.groq.com/openai/v1")
-_MODEL = os.getenv("LLM_MODEL", "llama-3.3-70b-versatile")
+# Provider presets: (OpenAI-compatible base URL, default model). Pick one with
+# LLM_PROVIDER; override the pieces individually with LLM_BASE_URL / LLM_MODEL.
+_PRESETS = {
+    "groq": ("https://api.groq.com/openai/v1", "llama-3.3-70b-versatile"),
+    "gemini": (
+        "https://generativelanguage.googleapis.com/v1beta/openai/",
+        "gemini-2.0-flash",
+    ),
+    "openrouter": (
+        "https://openrouter.ai/api/v1",
+        "meta-llama/llama-3.3-70b-instruct:free",
+    ),
+}
+
+_provider = os.getenv("LLM_PROVIDER", "groq").lower()
+_preset_base, _preset_model = _PRESETS.get(_provider, _PRESETS["groq"])
+_BASE_URL = os.getenv("LLM_BASE_URL", _preset_base)
+_MODEL = os.getenv("LLM_MODEL", _preset_model)
 
 _SYSTEM = (
     "You are DocuAsk, a document question-answering assistant. Answer the "