Generative AI topics

Can you provide a list of Ai Chatbots that Specialize in handling government forms and applications?

frankscharley3 — Fri, 06 Nov 2026 23:37:21 GMT

Yes. Several AI-powered platforms are worth exploring if your goal is to simplify government forms, permit applications, and resident service requests.

Some options to consider include:

GovAssist by App Maisters Government: Supports government service workflows involving permits, licensing, benefits, taxes, and resident inquiries.
Vulcan Trajan: Helps residents identify the correct government process, understand requirements, complete forms, and navigate permits and registrations.
TechForGov IGNA: Combines AI assistance with digital forms, permit applications, notifications, and case management.
Ordinal: Offers AI-powered assistance for local governments, helping residents find information through approved government documents and helping staff with permitting-related information.

When evaluating these solutions, prioritize features such as application guidance, document collection, eligibility checks, missing-information alerts, application status updates, and integration with existing government systems. These capabilities can help reduce incomplete submissions, repeated inquiries, and manual administrative work.

403: Permission Denied for certain foundational models

gautamrjt — Wed, 04 Nov 2026 19:08:53 GMT

Hello,

I have an individual account (Azure Databricks, Premier, East US)without any support contracts, and I am using it solely for personal projects. I am getting this error whenever I am trying to use the GPT and/or Claude models (system.ai.claude-sonnet-5-5). However, I am able to use models like "system.ai.meta-llama-3-3-70b-instruct". I am getting this error for the models mentioned. I have enabled Unity Gateway. How do I overcome this?

403 PERMISSION_DENIED

Access to this model is disabled for your workspace.
Your Databricks account may require verification
or additional model-access approval.

Contact Databricks Support or your Databricks
account team for assistance.

Thanks.

Gautam Radhakrishnan Ajit

How to size your Genie bill (Genie Agent, Genie Code) before the free period ends (31 January 2027)

Valeria_Koz_DBX — Wed, 04 Nov 2026 16:46:12 GMT

31 January 2027 is closer than it looks. Until then, every question your people ask Genie One or a Genie Agent is free. From 1 February, it's billed like the rest of Genie: a free monthly allowance per person, then pay-as-you-go.

If nobody has worked out the number by then, finance will find it on the first invoice. And the conversation won't be about rolling Genie out further. It will be about who switched it on.

There's also less time than the date suggests. A usage pilot ideally needs 30 days over a normal period (or a min of two weeks), with Christmas sitting in the middle, and you'll want a couple of weeks to agree limits with finance. Count back and you need to start by mid-December. Here's how to get a number you can defend.

First, what's free and what isn't

Free until 31 Jan: Genie One and Genie Agents usage by people. During the promotion it doesn't use up the free allowance and doesn't count towards budgets.
Free every month: 150 DBUs per person, reset on the 1st.
Already billed today: Genie Code beyond the 150 DBUs, and everything a service principal does. Service principals get no free allowance and aren't covered by the promotion.

All free usage is recorded under the GENIE_FREE_USAGE SKU in system.billing.usage (from 20 July 2026). That's what makes it possible to price it.

Step-by-step

STEP 1. Turn on an alert-only budget now. In the account console, go to Usage → Budgets → Create budget. Choose the Unity Gateway resource type and the databricks-product: genie tag, and don't add other tags. Add a shared threshold with only Send alert ticked, plus a lower one at about 80% as an early warning. Don't block anyone yet. Today this budget only tracks billed usage (Genie Code and service principals), but it's the safety net you'll build on in step 6.
STEP 2. Pick one representative team and run a 30-day pilot. Fourteen days is the minimum, and 30 is better. Choose a team that uses Genie for real work, not your most enthusiastic early adopters. Their usage is what you'll scale up from.
STEP 3. Measure what the pilot would have cost. Run this on the pilot period, or on last month if people already use Genie. It adds up every Genie DBU per person, free and billed, then subtracts the 150 DBU allowance:
1. Keep the pilot inside one calendar month. The allowance resets on the 1st, so a pilot that spans two months needs this query run once per month.
STEP 4. Turn DBUs into a range, not a single number. The Monitor and understand your Genie cost page has an "equivalent cost of free usage" query that looks up your price per DBU and applies today's 25% promotional discount. Take list_price_per_unit from its output and give finance two numbers: billable_dbus × price × 0.75 and billable_dbus × price. The docs don't say whether the discount continues after the promotion, so the range is the honest answer.
STEP 5. Scale it up, then split people into three tiers. Multiply by the number of people you expect to use Genie in February, not by everyone with access. Then group them the way the Genie consumption guide suggests:
- Beginners: occasional use. Most stay within the free allowance.
- Practitioners: regular use. Set their limit from your median and 75th percentile.
- Power users: long-running agent work. Highest limit, given as a group override.
- Create one account group per tier, so limits follow people rather than names on a spreadsheet.
STEP 6. Set per-person limits before 1 February. Edit the budget from step 1. Keep the shared thresholds as alert-only, and add a per-user threshold for everyone (the beginner level), plus group overrides for practitioners and power users. Use Block usage on the per-user limits, not the shared one, so one heavy user can't block everyone.
STEP 7. Check again in February. Run the step 3 query on the first full billed month and compare it with your estimate. Move people between tiers where the data says so.

What good looks like (expected outcome)

Finance has a monthly range for Genie before the promotion ends, with the query behind it.
Every person sits in a tier, and each tier has a limit based on real usage, not guesses.
One person's heavy month triggers an alert or a personal block, never a company-wide one.
The first February invoice falls within the range you gave.

Common errors (and how to avoid them)

Reading the budget page as the bill. During the promotion, Genie One and Genie Agents usage doesn't appear against budgets at all, and the page resets every month. Size from system.billing.usage.
Forgetting service principals. An app or integration running on one shared service principal is billed from day one, with no allowance, and every user's questions land on that one identity. Check for these first.
Forgetting the SQL warehouse. Genie budgets track only the LLM usage. The warehouse that runs Genie's SQL is billed separately, so size it too (I covered warehouse budgets here).
Subtracting the allowance twice. The query in step 3 subtracts it once from total usage. If you work from billed spend only, the allowance is already taken out.
Setting beginner limits too low. A limit that blocks someone in their first week stops them before they've found the value.
Starting in January. A two-week pilot over the new year won't show you a normal month.

The takeaway

The free period isn't the time to stop measuring. It's the cheapest time you'll get to find out what Genie really costs, so finance hears a number from you before they see it on an invoice.

Do non-ASCII characters in file names break document indexing over a Unity Catalog volume?

suranga — Fri, 30 Oct 2026 23:38:10 GMT

I spent a few days on a document retrieval problem that turned out to have nothing to do with the retrieval setup, and I want to check whether anyone else has hit the same thing.

Setup: around 40 internal PDF documents in a managed Unity Catalog volume, indexed for question answering. Ordinary text PDFs, not scans. I confirmed that by pulling text out of one of them and getting roughly 77,000 characters.

Symptom: the index reported every file processed with zero errors. Every question came back with nothing retrieved. Not a wrong answer, not a partial one. Empty, and no error anywhere in the chain.

Two things came out of this and I think they are worth separating.

1. "Every file processed, 0 errors" was true and useless

That count measures files enumerated, not text landed. It cannot tell a real extraction apart from a silent empty one. So for several days I had a green status and no working retrieval, with nothing in between to tell me which I had.

What I would want instead is a per-document character or chunk count, and an explicit count of files that indexed to zero retrievable content. If your only health signal cannot distinguish success from silence, it is not a health signal. That holds for any parse-then-index pipeline over a volume, so if you are building one, add the output-side count before you need it rather than after.

2. File names, and this is the part I am actually asking about

The catalog, schema and volume names were all lowercase letters, digits and underscores. Clean ASCII. The file names were not. Six of the documents had Swedish characters in the name, and every file name contained spaces. Names along these lines:

Lönepolicy 1995.pdf

Jämställdhetspolicy 3054.pdf

Policy för hybridarbete 57.pdf

ASCII is 95 printable characters: the Latin letters, the digits, the space and common punctuation. Everything carrying a mark, so every Swedish, German, French, Polish or Czech letter, sits outside it. Most of a pipeline is character agnostic right up to the one hop where a file name becomes part of a path, a URL, a JSON key or a shell argument, and that is where I would expect this class of bug to live.

So I copied the volume, kept only the documents whose names were plain ASCII, and deliberately kept the spaces, so that non-ASCII characters were the only variable that changed. Same compute, same documents, same configuration.

Being straight about the limits of that: in the run where answers came back correctly, I still had the original volume and the ASCII copy attached at the same time, so an answer could have come from either one. I am rerunning it against the ASCII copy alone before I claim the file names are the cause. I will post the result in this thread either way.

What I am hoping someone can tell me:

Has anyone seen non-ASCII characters in file names break document parsing or indexing over a UC volume, silently or otherwise?
If you work with Nordic, German or Polish documents, do you normalise file names to ASCII on the way in as a matter of course?
Once a file name has been flattened, where do you keep the readable original title? A metadata column alongside the parsed text, or something else?

ai_decide() failing with Endpoint is not enabled for batch inference in Free Edition

data_pulse — Fri, 30 Oct 2026 15:35:41 GMT

I'm running ai_decide() in a notebook (SQL cell, serverless compute) and it's now consistently failing with:

[INTERNAL_ERROR] Remote session creation failed: INVALID_ARGUMENT: BAD_REQUEST: Endpoint ai_decide is not enabled for batch inference in this workspace. SQLSTATE: XX000

Details:

Workspace is on the Free Edition, AWS.
ai_decide() function was available and ran successfully in early October in Free Edition.
Even a minimal single-row ai_decide() test fails, so it's not query-specific.
Partner-powered AI features are enabled in workspace settings, so that's not the blocker.

Questions:

Is this expected? Is ai_decide(Beta) no longer available for batch inference on Free Edition workspaces, or was it a time-limited preview?
If it's still supported, what else needs to be enabled: is there a separate toggle or enrolment ?

How to enable DeepSeek V4.1 Flash (1017) on a new AWS workspace?

sillytail_dev — Wed, 28 Oct 2026 15:41:15 GMT

Databricks Support referred me to the Community for help with account verification and model access.

I have a newly created Databricks on AWS workspace. The account console shows Premium with AWS Marketplace billing configured. I authenticate using the official Databricks CLI as a workspace administrator.

The exact model system.ai.deepseek-v4-1-flash-1017 is listed in Unity Gateway model services, and my administrator already inherits EXECUTE on system.ai.

A minimal synthetic request to POST /ai-gateway/mlflow/v1/chat/completions still returns HTTP 403 PERMISSION_DENIED:

"Access to this model is disabled for your workspace. Your Databricks account may require verification or additional model-access approval. Contact Databricks Support or your Databricks account team for assistance."

Could a community engineer clarify the official verification and approval process for this exact model on AWS, including any region or trial-account restrictions? Can existing trial credits be used once access is approved?

Help Center SSO succeeds, but the resulting page says this account has no Help Center access. Which official private channel or team can verify the account and enable model access? I can provide account and workspace identifiers privately to Databricks staff.

Thank you.

Google Shopping Feed Optimization for Shopify Stores

UnfoldCRO2 — Wed, 28 Oct 2026 10:31:11 GMT

our Google Shopping feed, not your bids, is the real growth lever most Shopify stores ignore. Here’s how to optimize product data for Merchant Center, Shopping and Performance Max.

Most Shopify stores spend their Google Ads effort on bids, budgets and creative, while the product feed goes to Google exactly as the store exported it. That’s backwards. Shopping and Performance Max campaigns have no keyword list. Google reads your feed to decide when a product shows, who sees it, and how the listing is described. If the feed is weak, bid strategy can only do so much. If the feed is clean, every rupee of spend has better material to work with.

How your product data reaches Merchant Center

On Shopify, product data reaches Google Merchant Center in one of two ways:

The native Google & YouTube app. It syncs your catalog automatically, mapping Shopify fields to Google attributes on a schedule.
A dedicated feed app or a direct API integration. Larger or more complex catalogs often use one of these for finer control over field mapping, exclusions and extra data.

Either way, Merchant Center sits in the middle. It validates every item, flags problems, and passes approved products to your campaigns.

The key thing to understand is that Shopify fields were built for your storefront, not for Google. Your titles, descriptions, tags and variants were written for people browsing your site. Feed optimization means reshaping that data into what Google’s matching and ad surfaces actually reward.

Write titles that match how people search

The title is the highest-impact field in the feed, because Google matches search queries against it heavily. A storefront title like “Aria Midi Dress” might look clean on your site, but it misses most of the searches that should trigger it.

Build titles from the terms shoppers actually type, most important first. A pattern that works for most catalogs is Brand + Product Type + Key Attribute + Color + Size or Material. For apparel, that could be “Zoya Cotton Midi Dress, Floral Print, Navy, Size M”. Google allows up to 150 characters, but listings show far fewer, so put the words that decide the click within roughly the first 70.

Keep descriptions factual

Descriptions carry less weight than titles, but they still feed into matching. Make the first sentence cover what the product is, what it’s for, and its standout attributes, then add detail.

Leave out store policies, shipping terms and promotional language. They add nothing to matching and can trigger policy flags.

Images that pass review

Google expects a clean primary image, usually on a white or plain background, with no promotional overlays, watermarks or added text. Overlays such as “SALE” badges are one of the most common causes of disapproval.

Use your highest-resolution product shot as the main image, and supply additional images so Google has options to rotate. Many of the image SEO basics from organic search carry straight over to the feed: file quality, sizing and naming.

Identifiers: GTIN, MPN and identifier_exists

Unique product identifiers tell Google exactly what you sell. They’re required for most branded and manufactured goods.

gtin: supply the barcode number wherever you have one.
mpn: for products with a manufacturer part number, supply it together with brand.
identifier_exists: for custom or handmade products with no barcode, set this to no. Don’t leave the fields blank, and never invent a value.

Correct identifiers improve matching. Some listing features are only available to products that have them.

Two category fields, two different jobs

google_product_category maps each item to Google’s official taxonomy. Set it as deep as the taxonomy allows. Stopping at a broad top-level node gives Google less to work with.

product_type is your own category string. Mirror your real site structure here, for example “Home > Kitchen > Cookware > Non-stick Pans”. It becomes useful when you split campaigns, listing groups and reports by category.

Keep price and availability in sync

Google regularly checks the price and availability in your feed against what it crawls on the landing page. A mismatch leads to disapproval. Common causes are:

A sale price that hasn’t reached the feed
The wrong currency
An out-of-stock variant still marked as in stock

Check that the sync pushes sale_price correctly and that availability reflects real inventory. For stores selling in India, confirm that the currency and the tax-inclusive price in the feed match what the live product page shows.

Variants and custom labels

Submit each variant as its own item, and give variants of the same product a shared item_group_id. Google can then group colors and sizes together while still advertising the exact option a shopper searched for. Each variant needs its own identifier, image and availability.

Custom labels are never shown to shoppers. They exist to help you structure campaigns. You get five fields, custom_label_0 to custom_label_4. Use them to tag products by:

Margin
Bestseller status
Season
Price band
Clearance

Then build campaigns and budgets around those tags. This is how you stop treating every SKU the same and move spend toward the products that actually make money.

Disapprovals and supplemental feeds

Check the Diagnostics section in Merchant Center regularly. A disapproved item can’t serve at all. The usual causes are:

Images: overlays, low resolution, or a missing primary image
Price and availability: mismatches between the feed and the landing page
Identifiers: missing or invalid values
Policy: violations in titles or descriptions
Landing pages: pages that fail review or don’t load correctly

Sometimes you can’t, or don’t want to, change the data at source in Shopify. In that case, use a supplemental feed to override or add specific fields, such as better titles, custom labels or corrected categories.

A supplemental feed matches rows to your primary feed by product ID. With the Google & YouTube app, those IDs follow the app’s own format, which combines the country code, the product ID and the variant ID. Export the IDs from Merchant Center before building the sheet, so every row matches. This lets you layer optimization on top of the automated sync without fighting the app’s field mapping.

Why the feed drives campaign performance

Shopping and Performance Max both pull their inventory, relevance signals and product copy straight from the feed. Google’s automation can only work with the data you give it:

Clear titles
Correct identifiers
Accurate categories
Useful custom labels

With that data in place, the automation has better inputs, and better inputs usually mean more relevant impressions at a lower cost. Strong organic collection pages help as well, since paid and organic listings for the same products reinforce each other.

How UnfoldCRO helps

We audit your Shopify-to-Merchant-Center sync, rewrite titles and categories at scale, fix disapprovals, add identifiers and custom labels, and build supplemental feeds for Indian D2C and ecommerce brands running Shopping and Performance Max. Feed work isn’t glamorous, but it’s where most of the easy wins are.

Needs verification

API name. Your draft said the Google & YouTube app syncs “over the Content API”. Google has been moving merchants from the Content API for Shopping to the newer Merchant API, so I left the API name out rather than risk naming a deprecated one. Confirm which applies before you add it back.
Missing links. Your draft mentioned the image SEO guide and the Shopify collection page SEO guide but had no URLs, so I kept them as plain mentions. Send me the URLs if you want them linked.
New detail. The note on supplemental feed IDs (country code, product ID and variant ID) is standard behaviour for the Google & YouTube app. It wasn’t in your original, so remove it if you’d rather not add new technical detail.

Best Architecture for Building a Custom Agent in Databricks?

Vamsi3757 — Tue, 20 Oct 2026 21:57:44 GMT

Hi Everyone,

I'm building a Databricks cost optimization agent and would appreciate some guidance on the recommended architecture.

My current design consists of the following functions:

Retrieve the last N job runs along with cost, DBU usage, runtime, cluster configuration, and worker information.
Read and analyze Spark event logs from Compute Log Delivery stored in a Unity Catalog Volume.
Combine historical run metrics and Spark event metrics into a consolidated dataset for analysis.

My goal is for a user to ask questions such as:

"Analyze job_id <job_id> and recommend cost optimizations."

The agent would analyze up to the last 50 runs (or however many runs have logs available), evaluate cost, utilization, shuffle, spills, executor usage, memory, etc., and then generate optimization recommendations.

My functions are currently implemented as Python functions and are not Unity Catalog Functions. Given this use case:

Is it better to build this using Agent Framework (Code Your Own Agent)?
Should I convert some or all of these functions into UC Functions?
Would a Genie Space, Agent Framework, or Databricks Apps be the recommended architecture for this type of cost optimization agent?

I'm interested in understanding Databricks best practices for building an agent that needs to orchestrate Python logic, inspect Spark event logs, and generate recommendations using an LLM.

Thanks in advance for any recommendations or design suggestions.

Genie One Answers Likes and Dislikes Button

SenthilKrishna — Tue, 20 Oct 2026 13:59:39 GMT

Genie one answers shows an option to like or dislike answers. Where do they get stored? Is there a way for me to identify how many dislikes of the Q&A pair, and perform an analysis over that pattern and come up with proper instructions, benchmark testing or examples to respective agents?

Genie One Guardrails

SenthilKrishna — Tue, 20 Oct 2026 12:15:42 GMT

We are opening up the Genie AI to the business users. The Genie One seems to be a good fit, but the problem I noticed is the Guardrails and influence over the single orchestration. All the questions have been nicely interrupted by Genie one and works with agents, tables from all schemas, notebooks, etc. My issue here is that I want to prioritize the search to go on the agents first, Clarifying questions for underspecified prompts than assuming the metrics, categories and date ranges. I feel like that's possible by creating the persona via the skill under Genie One, and customize with guardrails. Is that a right approach?

Databricks Genie One - Managing past conversation initiated by Databricks service principal account

ken_npk — Sun, 18 Oct 2026 14:32:19 GMT

Hello everyone.

Recently, our team use Databricks service principal account, connecting to Genie One MCP server (via unity gateway) for data retrieval. However, we notice that there are currently no ways for us to manage the past conversations that had happened with service principal.

We not able to find SDKs or APIs to list all or perform deletion on past conversations initiated by service principal. As far as we know, there is no way for us to login as service principal and perform deletion via Genie One interface. We also noted that those functionalities only available for Genie Agent.

This limitation is problematic for us. In some cases, Genie One has misinterpreted our data during its reasoning process. Although we have corrected the relevant issues in the data semantic layer, Genie One continues to rely on past conversations containing incorrect interpretations when answering user questions. Since we are unable to delete these past conversations, it is difficult for us to prevent Genie One from reusing misleading context.

Are there any future plans on having a way for listing all or performing deletion on past conversations that were done by service principals? Or are there any alternative ways to manage the past conversations by service principals?

Appreciate any insights and suggestions from Databricks community. Thank you.

Per-agent guardrails: what do you alert on, tokens, retries, or cost per task?

DoTA — Sun, 18 Oct 2026 00:56:24 GMT

I'm trying to work out what a sensible guardrail looks like for each agent when several of them share a few serving endpoints.

A plain token budget feels too blunt. An agent that retries a failing tool four times looks like a heavy user, while a cheap agent stuck in a loop can burn far more than its neighbours before anyone notices.

The options I'm weighing:

1. A hard daily cap per agent on tokens, through AI Gateway rate limits.

2. An alert on tool invocations per request, to catch loops and retry storms early.

3. An alert on cost per completed task against that agent's own 7-day baseline.

For those running agents in production on Databricks: which of these do you actually use, and which one did you regret? And is anything in the system tables or inference tables good enough to build that baseline from, or do you end up logging your own per-task numbers?

Happy to hear partial answers too, even just what you stopped doing.

Monitoring a custom Agent (Responses) serving endpoint with an experiment on the Unity Catalog

FantineM — Mon, 12 Oct 2026 11:53:15 GMT

Hi,

I have a Langgraph agent deployed in Databricks on a custom serving endpoint of type Agent (Responses).
So far, I have been tracing this agent on a "classic" monitoring experiment. I was using an ad hoc job to sync those traces to a table in the Unity Catalog. But now, one can create an experiment with a location in the UC directly, which sounds handy for my purpose. However, after I redeploy the agent (I set the mlflow experiment to the new one with UC location before deploying and I set MLFLOW_EXPERIMENT_ID as en environment variable for the deployed entity) I cannot see any traces in the new experiment.
The only way I could ever make this work was creating a span from a notebook and running a non deployed agent from there.
Am I missing something? Has anyone succeeded to make it work who could help me out?
Thanks in advance for any help or suggestion 😉

Unable to Serve Gemini 3.8 Flash on Azure Databricks – East US

krSindhu — Fri, 09 Oct 2026 06:46:18 GMT

I am working with an Azure Databricks workspace in the East US region and am trying to use the Gemini 3.8 Flash model through Databricks Model Serving.

The model is visible in the workspace under the system AI models as:

system.ai.databricks-gemini-3-8-flash

However, when I try to create a serving endpoint, I receive the following error:

“Provisioned throughput is unavailable. This model does not support on-demand provisioned throughput. Reserved capacity is not currently available for this model in this workspace. Select a different model.”

I have attached a screenshot of the error for reference.

Could someone please clarify what is the recommended way to access it (pay-per-token/ADI Services or another approach) and what workspace-level configuration or enablement is required?

How do you regression-test a Genie space or agent when a table or metric view changes underneath it?

DoTA — Wed, 07 Oct 2026 14:43:28 GMT

I hope this is useful for anyone running Genie spaces or agents past the demo stage. This is a question more than a recommendation, because I have not found a pattern I fully trust yet.

The situation: a Genie space or an agent sits on top of a handful of tables or a Unity Catalog metric view. Someone renames a column, changes a join key, redefines a measure, or backfills a table, and the answers shift without any error. The SQL still runs. The number is just different, or the wrong table gets picked. Nothing in the usual pipeline tests catches that, because the pipeline itself is healthy.

What I am trying to work out:

1) What do you use as the regression set? Genie spaces have benchmark questions with reference SQL, and some teams keep a separate golden set outside the space (30 to 50 questions with a reviewed answer each). Do you keep both, and where does the set live (the space, Git, a table)?

2) What counts as a pass? Exact SQL match is too strict (equivalent queries differ). Result-set equality is better but breaks when data legitimately moves. Do you compare against a pinned snapshot, a tolerance on the numbers, or an LLM judge for the explanation text?

3) What triggers the run? On every deploy of the bundle, on a schedule, or when lineage says an upstream table or metric view of the space has changed? Has anyone wired the system.access lineage tables to decide which spaces to re-test?

4) Cost and speed: how long does your set take to run, and do you run the full set or a sample on each change?

5) Who owns a failing test, the data engineer who changed the table or the person who owns the space?

If you already have this running, even a rough description of the loop would help. And if you decided it was not worth building, I would like to hear why.

Supervisor Agent Visualisations

haaland — Wed, 07 Oct 2026 06:43:51 GMT

I created a supervisor agent and asked it to plot something.
It was able to generate text and table, then output a lot of gibberish which was supposed to be its attempt of plotting.

Some of it's output (redacted):

{'text/plain': '<Figure size 1400x1000 with 2 Axes>', 'image/png': 'iVBORw0KGgoAAAANSUhEUgAABW0AAAPdCAYAAADxjUr8AAAAOnRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjEwLjAsIGh0dHBzOi8vbWF0cGxvdGxpYi5vcmcvlHJYcgAAAAlwSFlzAAAPYQAAD2EBqD+naQABAABJREFUeJzs3Xd4U+UXwPFv0l3oppNZ9pIhs+wNgiiCskSmgAqi4MQF+HPi3ojIEAQFBGUIsvdGEMoeZXcCnXQlub8/Qi9NZ9ombVPO53l4aO58k5N7k5z73vNqFEVREEIIIYQQQgghhBBCCFEqaEu6AUIIIYQQQgghhBBCCCHukaStEEIIIYQQQgghhBBClCKStBVCCCGEEEIIIYQQQohSRJK2QgghhBBCCCGEEEIIUYpI

Is this a bug and will it be fixed?

Practical Generative AI Skills for Beginners and Working Professionals

asmorix — Wed, 07 Oct 2026 00:21:41 GMT

Hi everyone,

At Asmorix, we regularly interact with learners who are interested in moving into Generative AI and AI-related careers. One thing we’ve noticed is that many beginners are unsure where to start because the GenAI ecosystem is evolving so quickly.

A good starting point is to understand the fundamentals of Python, machine learning and data handling, followed by concepts such as LLMs, prompt engineering, RAG, vector databases and model evaluation.

We also encourage learners to work on practical projects rather than focusing only on theory. Building a simple RAG application, chatbot, document question-answering system or AI-powered data application can help develop a better understanding of how GenAI is used in real-world scenarios.

We’d love to hear from the Databricks community:

Which GenAI skills are most valuable for beginners today?
What projects would you recommend for building practical experience?
How important are Databricks and MLflow for someone starting a GenAI career?
What should learners focus on to become job-ready?

At Asmorix, we believe continuous learning and hands-on practice are key to building a strong foundation in emerging technologies.

Looking forward to learning from the community and sharing experiences.

Thanks!

Account Admin unable to configure Unity Gateway rate limits for system.ai model services

HariUmeshNaraya — Tue, 06 Oct 2026 16:59:06 GMT

I am working with Databricks Unity Gateway / Unity Catalog model services and am trying to configure rate limits for system-provided model services under system.ai.

##Issue

For example, when I open Unity Gateway, select system.ai.claude-sonnet-5, and try to configure the rate limits, the UI displays: "User does not have permission to access this resource." I see the same behavior for other system.ai models.

##Current permission

I have Account Admin access at the Databricks account level. However, when I run the following SQL command:

SHOW GRANTS ON SCHEMA system.ai;

the result only shows the following permissions for account users:

EXECUTE
READ VOLUME
SELECT
USE SCHEMA

There is no MANAGE privilege shown on system.ai.

Additional observation

When I directly call system.ai.claude-sonnet-5, I receive HTTP 403 with the following message:

PERMISSION_DENIED: The endpoint is temporarily disabled due to a Databricks-set rate limit of 0.

I am not sure whether this is related to the inability to configure the rate limit in the UI or whether it is a separate platform-side restriction.

Questions

1. Why does Account Admin access not automatically allow me to configure rate limits for system.ai model services?

2. Is Metastore Admin access required to configure rate limits for system-provided model services?

3. Can an Account Admin grant MANAGE on system.ai to themselves or another user/group, or must this be done by a Metastore Admin?

4. What is the least-privilege approach for allowing an AI platform administrator to manage Unity Gateway rate limits without giving them full Metastore Admin privileges?

5. Is MANAGE on the system.ai schema sufficient to configure rate limits on the system-provided model services?

6. Is the "Databricks-set rate limit of 0" error expected when a model service has not been enabled/configured, or does it indicate a separate Databricks-side restriction?

What I am trying to understand

I would like to clarify the distinction between account-level administration and Unity Catalog / metastore-level permissions for system.ai model services, and understand the correct least-privilege configuration needed to manage Unity Gateway rate limits.

Relevant SQL output

SHOW GRANTS ON SCHEMA system.ai;

Principal: account users
ActionType: EXECUTE
ObjectType: SCHEMA
ObjectKey: system.ai

Principal: account users
ActionType: READ VOLUME
ObjectType: SCHEMA
ObjectKey: system.ai

Principal: account users
ActionType: SELECT
ObjectType: SCHEMA
ObjectKey: system.ai

Principal: account users
ActionType: USE SCHEMA
ObjectType: SCHEMA
ObjectKey: system.ai

What Is – Agent mode in Genie? (Quick answers vs investigations)

Valeria_Koz_DBX — Tue, 06 Oct 2026 16:00:52 GMT

Imagine asking a colleague two questions on the same morning.

The first: "What was our travel spend last month?" They open a report, read off a number, and you're done in a minute. That's a lookup.

The second: "Why is our travel spend so much higher than last year?" A good colleague won't answer that with one number. They'll think of a few possible reasons, such as more trips, pricier trips or one team travelling more, check each against the data, and come back with a short write-up and the evidence. That's an investigation.

Genie, the Databricks tool that lets you ask questions about your data in plain English, can now work both ways. The second way is called Agent mode.

The two ways Genie can answer

Chat mode is the quick-answer colleague. Genie writes one query and shows you the result as a table or chart. It's best for "what", "how much" and "how many" questions.

Agent mode is the investigating colleague who asks "why". Genie plans its approach, runs several queries one after another (each informed by the last), and writes up what it found in a short report. You can follow every step and the query behind it, so you can check its reasoning instead of just trusting the conclusion.

A real-world example

A finance team I worked with kept facing the same question: "Why are we over budget this quarter?" Answering it meant an analyst pulling several reports into a spreadsheet and writing an email, which took days.

With Agent mode, a manager can ask directly. Genie breaks the overspend down by department, then by supplier, spots that most of it comes from two suppliers, and writes it up. The analyst's job shifts from building the answer to checking it.

When to use which

Use chat mode when you know exactly what you want: a number, a list, a chart.
Use Agent mode for "why" and "what's driving" questions. Agent mode is slower because it does more work, so it's the wrong choice for simple lookups.

One caution: an investigation is only as good as the briefing. If Genie hasn't been told how your team defines "spend" or "budget", you'll get a confident report about the wrong thing. Read the steps, not just the conclusion.

What's the first "why" question you'd hand to an investigating colleague?

What is the most effective way to evaluate Genie Agents/ Genie Spaces.

Yash-Verma — Tue, 06 Oct 2026 09:11:29 GMT

I would like to understand the most effective way to evaluate the Genie agent results. IS there any option we put as a guardrail for "Genie One" such that it doesn't go beyond the Genie spaces for searches apart from putting in prompt.