Check out some of the Samples!

[i2amsam]

Check our our coffee shop and menu samples showing a simple coffeeshop and then a more complex RAG workflow

[TAxdorff]

One of the things that could be helpful is a sample showing how you can use genkit inside of other firebase functions.

Problem:

• Currently all documentation for deploying to firebase shows using onFlow to create standalone functions that can be triggered from client SDKs through the callables implementation
• This doesn't cover implementing flows/prompts as part of background functions like pubSub bindings, or firestore document brindings

Having a couple samples using onMessagePublished, onDocumentUpdated (ideally with AuthContext) would go a long way to demonstrate how to retrofit existing projects a lot better then the current documentation.

My assumption is we should just use definePrompt / defineDotprompt, but is there a way to use Flows? Do you get the Telemetry benefit's if you only use prompts? Etc.

[pascalbe-dev]

I'm also very interested in that.
I believe, it's a quite common use case to run gen AI in the background when a business operation finished behind the scenes.

If you have any examples or hints, please share :)

Great work on genkit by the way 🥇

[peterfriese]

Great feedback for enhancing the documentation.

I'm working on a blog post and video that will show how to implement a RAG use case.

I will show how to use onCall so you can call your gen AI flow from a mobile app, but also how to use onDocumentCreated to call a flow when a new document is added to Firestore.

Here is a code snippet from the blog / video that highlights how to call a flow from a Firestore trigger:

export const updateSummary = onDocumentUpdated("articles/{documentId}", async (event) => {
  const newData = event.data?.after.data() ?? {};
  const previousData = event.data?.before.data() ?? {};

  if (!newData.fullText || newData.fullText === previousData.fullText) {
    console.log("fullText is empty, or there was no change.")
    return;
  }

  const text = newData.fullText

  const summary = await runFlow(summarizeTextFlow, text)

  await event.data?.after.ref.update({
    notes: summary
  });
});

[TAxdorff]

I've sorta gotten one to work, but it doesn't do so consistently and I notice two things:

• The generate seems to retry on it's own, as I don't see my error logs firing for catching errors or lack of data. But it's unclear why. I know the logger works as I see the logs firing in some of the lines cut from sample code below.
• Running a firestore document update in a run (from @genkit-ai/flow) never seems to work, hence the commented out copy below. Could be just the limited testing I did and the inconsistency of getting a result above.

Gotta say, I love the trace generation of GenKit, and that the plugin system seems well designed, and is ramping up quickly! Getting something that works with both Gemini and OpenAI is really useful (which Instructor has yet to implement). Keep up the good work!

Leaving the code below if it might help others see how you can integrate it similar to the code @peterfriese showed above

Sample Flow

const flowInputSchema = z.object({
    user_id: z.string(),
})
export type flowInput = z.infer<typeof flowInputSchema>;

const flowOutputSchema = z.object({
    attempts: z.number(),
    errorCode: z.string().optional(),
    errorMessage: z.string().optional(),
})
export type flowOutput = z.infer<typeof flowOutputSchema>;

const FUNCTION_LOG_SIGNATURE = "myFlow"

export const myFlow = defineFlow(
    {
        name: 'myFlow',
        inputSchema: flowInputSchema,
        outputSchema: flowOutputSchema,
    },

    async (input: flowInput) => {
        // Grab Firebase Document from user_id
        const [success, docData, docRef] = await run('grabDoc', async () => {
            return await grabDoc(input.user_id, FUNCTION_LOG_SIGNATURE);
        });
        if (!success) {
            throw new Error();
        }

        // Do your other processing using docData
        //...


        model = gemini15ProPreview;
        config = {
            temperature: 0.0,
            maxOutputTokens: 4096,
            }

        let attempts = 1;
        let errorCode = 'uknowwn';
        let errorMessage = 'unknown';
        while (attempts <= MAX_ATTEMPTS) {
            try {
                const generate_repsonse = await generate({
                    model: model,
                    candidates: 1,
                    prompt: [
                        { text: "Evaluate image" },
                        { media: {contentType: 'image/jpeg', url: imageContent} },
                    ],
                    config: config,
                    output: {schema: MyZodSchemaIWriteToFirestore}
                });
                const data = generate_repsonse.output();
                if (!data) {
                    attempts += 1;
                    const message = FUNCTION_LOG_SIGNATURE + ' : ERROR : error loading response';
                    myLogger.error(message, {response: generate_repsonse});
                    continue;
                }
                // We've gotten a good output, lets update firestore and return reults
                // await run('updateDoc', async () => {
                //     await docRef.update({
                //         "debug.summary_model": model.name,
                //         "summary": data,
                //         "debug.llm_attempts": attempts,
                //     });
                // });
                await docRef.update({
                    "debug.summary_model": model.name,
                    "summary": data,
                    "debug.llm_attempts": attempts,
                });
                return {
                    attempts: attempts,
                };
            } catch(error) {
                attempts +=1;
                const e = error as GenkitError;
                errorCode = e.status;
                errorMessage = e.message;
                const message = FUNCTION_LOG_SIGNATURE + ' : ERROR : error with llm request';
                myLogger.error(message, {error: error});
                continue;
            }
        }
        // All attempts failed
        return {
            errorCode: errorCode,
            errorMessage: errorMessage,
            attempts: attempts
        }
    }
)

Sample Function:

exports.evalImage = onMessagePublished(funcConfig, async (event) => {
    // Load event data
    const user_id = event.data.message.json.user_id;

    myLogger.debug(`${FUNCTION_LOG_SIGNATURE} : DEBUG : Started myFlow for User: ${user_id}`)

    await runFlow(summaryFlow, {user_id});
    // Use the attempts / errorCode / errorMessage if necessary

    myLogger.debug(`${FUNCTION_LOG_SIGNATURE} : DEBUG : Finished myFlow for User: ${user_id}`)
});