Add Flask API which wraps the existing LLM based Python functionality

firebase.json

@@ -4,12 +4,23 @@
     "cleanUrls": true,
     "ignore": ["firebase.json", "**/.*", "**/node_modules/**"]
   },
-  "functions": {
-    "predeploy": ["yarn build:functions"],
-    "source": "functions",
-    "runtime": "nodejs18",
-    "runtimeConfig": ".runtimeconfig.json"
-  },
+  "functions": [
+    {
+      "predeploy": ["yarn build:functions"],
+      "source": "functions",
+      "codebase": "maple",
+      "runtime": "nodejs18",
+      "runtimeConfig": ".runtimeconfig.json"
+    },
+    {
+      "predeploy": [
+        ". llm/venv/bin/activate && python3 -m pip install -r llm/requirements.txt"
+      ],
+      "source": "llm",
+      "codebase": "maple-llm",
+      "runtime": "python311"
+    }
+  ],
   "firestore": {
     "rules": "firestore.rules",
     "indexes": "firestore.indexes.json"

infra/Dockerfile.firebase

@@ -1,7 +1,7 @@
 FROM andreysenov/firebase-tools:latest-node-18
 
 USER root
-RUN apt update && apt install -y curl
+RUN apt update && apt install -y curl python3 python3-pip python3-venv
 
 WORKDIR /app
 RUN chown -R node:node .

infra/firebase.compose.json

@@ -4,11 +4,22 @@
     "cleanUrls": true,
     "ignore": ["firebase.json", "**/.*", "**/node_modules/**"]
   },
-  "functions": {
-    "predeploy": ["yarn build:functions"],
-    "source": "functions",
-    "runtime": "nodejs18"
-  },
+  "functions": [
+    {
+      "predeploy": ["yarn build:functions"],
+      "source": "functions",
+      "codebase": "maple",
+      "runtime": "nodejs18"
+    },
+    {
+      "predeploy": [
+        ". llm/venv/bin/activate && python3 -m pip install -r llm/requirements.txt"
+      ],
+      "source": "llm",
+      "codebase": "maple-llm",
+      "runtime": "python311"
+    }
+  ],
   "firestore": {
     "rules": "firestore.rules",
     "indexes": "firestore.indexes.json"

llm/.gitignore

@@ -0,0 +1,4 @@
+venv/
+__pycache__/
+databases/
+.secret.local

llm/llm_functions.py

@@ -278,6 +278,7 @@ def set_my_llm_cache(cache_file: Path=LLM_CACHE) -> SQLiteCache:
     Set an LLM cache, which allows for previously executed completions to be
     loaded from disk instead of repeatedly queried.
     """
+    cache_file.parent.mkdir(exist_ok=True)
     set_llm_cache(SQLiteCache(database_path = cache_file))
 
 @dataclass()

llm/main.py

@@ -0,0 +1,73 @@
+from flask import Flask, jsonify, abort, request
+from llm_functions import get_summary_api_function, get_tags_api_function
+import json
+from firebase_admin import initialize_app
+from firebase_functions import https_fn, options
+import os
+
+initialize_app()
+app = Flask(__name__)
+
+
+def is_intersection(keys, required_keys):
+    return (keys & required_keys) == required_keys
+
+
+def set_openai_api_key():
+    match os.environ.get("MAPLE_DEV"):
+        case "prod":
+            if os.environ.get("OPENAI_PROD") != None:
+                os.environ["OPENAI_API_KEY"] = os.environ["OPENAI_PROD"]
+        case _:  # if "dev" or unspecified, use OPENAI_DEV
+            if os.environ.get("OPENAI_DEV") != None:
+                os.environ["OPENAI_API_KEY"] = os.environ["OPENAI_DEV"]
+
+
+@app.route("/summary", methods=["POST"])
+def summary():
+    set_openai_api_key()
+    body = json.loads(request.data)
+    # We require bill_id, bill_title, bill_text to exist as keys in the POST
+    if not is_intersection(body.keys(), {"bill_id", "bill_title", "bill_text"}):
+        abort(404, description="requires bill_id, bill_title, and bill_text")
+
+    summary = get_summary_api_function(
+        body["bill_id"], body["bill_title"], body["bill_text"]
+    )
+
+    if summary["status"] in [-1, -2]:
+        abort(500, description="Unable to generate summary")
+
+    return jsonify(summary["summary"])
+
+
+@app.route("/tags", methods=["POST"])
+def tags():
+    set_openai_api_key()
+    body = json.loads(request.data)
+    # We require bill_id, bill_title, bill_text to exist as keys in the POST
+    # Note: & is essentially set intersection
+    if not is_intersection(body.keys(), {"bill_id", "bill_title", "bill_text"}):
+        abort(404, description="requires bill_id, bill_title, and bill_text")
+
+    tags = get_tags_api_function(body["bill_id"], body["bill_title"], body["bill_text"])
+
+    if tags["status"] in [-1, -2]:
+        abort(500, description="Unable to generate tags")
+
+    return jsonify(tags["tags"])
+
+
+@app.route("/ready", methods=["GET"])
+def ready():
+    return ""
+
+
+@https_fn.on_request(
+    secrets=["OPENAI_DEV", "OPENAI_PROD"],
+    timeout_sec=300,
+    memory=options.MemoryOption.GB_1,
+)
+def httpsflaskexample(req: https_fn.Request) -> https_fn.Response:
+    with app.request_context(req.environ):
+        return app.full_dispatch_request()

llm/readme.md

@@ -122,3 +122,74 @@ This project uses OpenAI's API for various language processing tasks. To use the
    ```python
    import os
    print(os.environ.get('OPENAI_API_KEY'))
+
+# Running the API
+
+Set up a virtual environment and run the Flask app
+
+```
+python3 -m venv venv
+source venv/bin/activate # .fish if using fish
+pip3 install -r requirements.txt
+python3 -m flask --app main run
+```
+
+## Infrastructure notes
+
+As of 2025-06-17, the version of `python3` inside the
+`infra/Dockerfile.firebase` is 3.11. Therefore, the `firebase.json` files use
+the `python311` runtime.
+
+## Deploying locally
+
+This is quite tricky due to how we overlay our current source directory to
+`/app` inside the container. You'll need to create and install dependencies from
+**inside** the container. If you are just working on python related code that
+doesn't need to be in Firebase, you **won't** be able to use this environment.
+
+```shell
+# Build the maple-firebase container
+yarn dev:update
+# Start up bash within the maple-firebase container
+docker run -v .:/app -it maple-firebase /bin/bash
+# Build the virtual env and install the dependencies matching the container
+python3 -m venv llm/venv
+source llm/venv/bin/activate
+pip3 install -r llm/requirements.txt
+```
+
+Note: you'll need to set `OPENAI_DEV` and `OPENAI_PROD` in a
+`llm/.secret.local` file. Get it with `firebase functions:secrets:access
+OPENAI_DEV`. They can be set to the same token. You can see the function URL
+in the emulator after running `yarn dev:up`.
+
+## Deploying to Firebase
+
+```shell
+# not sure if the GOOGLE_APPLICATION_CREDENTIALS is strictly necessary, but I
+# had a number of problems with authorization
+GOOGLE_APPLICATION_CREDENTIALS=/path/to/application_default_credentials.json \
+  firebase deploy --only functions:maple-llm --debug
+
+# Hit the function in production
+curl \
+  -X POST \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
+  -d '{"bill_id": "1234","bill_title": "A title","bill_text": "Some bill text"}' \
+  https://httpsflaskexample-ke6znoupgq-uc.a.run.app/summary
+```
+
+## Future work
+
+Currently, we are just using the built-in Flask server. We should switch to a
+production WSGI server, like `gunicorn`.
+
+The local emulator installation process is quite cumbersome and ideally the
+virtual environment was built during container instantiation instead of from
+within the Docker container (i.e. the "Deploying locally" docs above).
+
+The current API is a little wonky because we take `bill_id` **and** `bill_text`.
+We could just look up the `bill_text` via the `bill_id` using the Firestore API.
+It might make sense to avoid the HTTP wrapper all-together and figure out how
+JS <-> Python communication works without an HTTP layer.