Appwrite support chatbot

Harbor Help answers questions using 40 help articles and FAQs about a fictional project-management product. It retrieves excerpts from Appwrite VectorsDB, generates answers through OpenRouter inside an Appwrite Function, and links to the original help pages. The React interface runs on Appwrite Sites.

Requirements

• Node.js 22 or later and pnpm 10.
• An Appwrite project with VectorsDB, Functions, and Sites.
• An OpenRouter API key with credit for openai/gpt-5.6-luna, or another model that supports structured outputs.

Create the knowledge base

In your project on https://appwrite.io, create a VectorsDB database with ID support. Create a collection named Articles with ID articles, selecting the all-minilm embedding model, which uses 384 dimensions. Leave collection permissions empty; the Function reads with its dynamic API key.

Create a seeding API key with these scopes:

• vectorsdb.collections.read
• vectorsdb.indexes.read
• vectorsdb.indexes.write
• vectorsdb.documents.read
• vectorsdb.documents.write
• embeddings.write

git clone https://github.com/appwrite-community/support-chatbot.git
cd support-chatbot
pnpm install
cp .env.example .env

Fill .env with the project's endpoint, ID, and seed key. The endpoint is shown in Project settings. Run:

pnpm seed

The script creates and waits for an HNSW cosine index, generates embeddings, and writes the documents. IDs are deterministic, so rerunning the same seed updates existing documents. If you remove an article or shorten one enough to remove chunks, delete its obsolete documents separately. data/articles.json contains the demo's source content, including all prices and policies. These describe Harbor, not Appwrite.

Deploy the Function

Create a Node.js 22 Function named Support chatbot, with ID support-chat.

Setting	Value
Root directory when using Git	functions/chat
Entrypoint	src/main.js
Build command	npm install --omit=dev
Dynamic API key scopes	embeddings.write, vectorsdb.documents.read
Execute access	Any, for this public help-center demo
Timeout	60 seconds

Configure these Function variables before deployment:

Variable	Value	Secret
APPWRITE_DATABASE_ID	support	No
APPWRITE_COLLECTION_ID	articles	No
OPENROUTER_API_KEY	Your provider key	Yes
OPENROUTER_MODEL	openai/gpt-5.6-luna	No

Deploy the Function through GitHub or the Appwrite CLI.

For GitHub, fork the companion repository and connect your fork to the Function under Settings > Configuration > Git settings. Set the production branch to main and the root directory to functions/chat. Use src/main.js as the entrypoint and npm install --omit=dev as the build command. Pushing a commit to main creates, builds, and activates a deployment. See deploying Functions from Git.

For the CLI, install the Appwrite CLI and set projectId and endpoint in the supplied appwrite.config.json. From the project directory, run:

appwrite login
appwrite push functions --function-id support-chat --activate

The runtime supplies APPWRITE_FUNCTION_API_ENDPOINT, APPWRITE_FUNCTION_PROJECT_ID, and the x-appwrite-key request header. Do not copy the seeding key into the Function.

Run the interface locally

Set the three VITE_ values in .env. These are public configuration; never put an API key in a VITE_ variable. Add a Web app with hostname localhost under your project's Apps section.

pnpm dev

The browser calls Functions.createExecution and displays its answer and verified source links. Conversation history stays in React state and resets on reload. The last six messages are sent to help interpret follow-up questions.

Deploy on Appwrite Sites

Create a Site using Vite. Connect a fork of this repository, with the root directory left at the repository root. Use pnpm install for installation, pnpm build for the build, and dist for the output directory. Choose static rendering and set the fallback file to index.html so help article links work on refresh.

Set VITE_APPWRITE_ENDPOINT, VITE_APPWRITE_PROJECT_ID, and VITE_APPWRITE_FUNCTION_ID as Site variables before building. Choose an available Appwrite subdomain. Add that hostname as a Web app in the project if it is not already registered.

Push changes to the connected production branch to deploy the Site. The Function and Site are deployed separately. Update the vector documents with pnpm seed and redeploy the Site after changing articles so the source pages match the retrieved text.

Verification

pnpm test
pnpm build
pnpm test:live

The live suite requires a deployed Function and a local test key with executions.write in addition to the seed scopes. It checks guest access, paraphrased CSV export, a contextual follow-up, an exact price, an unsupported question, an instruction-injection attempt, and invalid input. LLM outputs can vary; a passing suite is evidence for these cases, not a guarantee against hallucination.

The Function retrieves six candidates. Similarity is retrieval evidence, not confidence. It asks the model to refuse unsupported answers and rejects nonexistent citation IDs. It does not execute model-generated code or URLs. Public execute access is suitable for this demonstration; production deployments should apply authentication and request limits appropriate to their audience and provider budget.

Repository structure

• functions/chat/src/main.js: request validation, retrieval, follow-up rewriting, and answer generation.
• functions/chat/src/knowledge.js: chunking and citation validation.
• scripts/seed.mjs: repeatable knowledge-base import.
• scripts/test-live.mjs: deployed Function checks.
• src/: Vite React chat and source article pages.
• data/articles.json: 40 fictional help articles and FAQs.

MIT licensed.