[Bug]: No automation YAML was returned.

[afreemo-glitch]

Home Assistant version

2026.5.4

AI Automation Suggester version

1.5.6

Provider and model

gemini-2.5-flash

What happened?

I have never been able to get your plugin working.
not sure if I am being a bit daft

Set everything up, reload the plugin and in Notifications i get this as the suggestion

_AI automation suggestion
The provider returned structured output that could not be parsed. Try regenerating with a lower entity limit or a newer model.

Warnings:

No automation YAML was returned._

Any suggestions?

Relevant logs

Service call data

[ITSpecialist111]

Thanks for the report — sorry it''s been frustrating to get going.

The message you''re seeing ("structured output that could not be parsed" + "No automation YAML was returned") means the integration reached the provider and got a response back, but couldn''t extract valid suggestion JSON/YAML from it. The parser already has several fallback layers, so to fix the actual cause I need to see what gemini-2.5-flash is returning.

Could you please grab debug logs? Add this to configuration.yaml, restart, reproduce once, then share the log:

logger:
  default: warning
  logs:
    custom_components.ai_automation_suggester: debug

Then Settings → System → Logs, filter by suggester, and paste the entries around the failed generation (redact any API keys). The raw provider response in those logs is what I need.

A couple of quick things to try in the meantime:

• Lower the entity limit in the integration options and regenerate.
• Make sure you''re on v1.5.7 (it includes a Gemini schema fix from 1.5.6).

Keeping this open until we see the logs.

[afreemo-glitch]

Hi there

I have installed 1.5.7 and re ran it

I am not getting this

Logger: homeassistant.components.recorder.db_schema
Source: components/recorder/db_schema.py:581
Integration: Recorder (documentation, issues)
First occurred: 22:13:19 (1 occurrence)
Last logged: 22:13:19

State attributes for sensor.ai_automation_suggester_google_ai_automation_suggestions_google exceed maximum size of 16384 bytes. This can cause database performance issues; Attributes will not be stored

[JamesHannon]

Also ran into the with 200 entities, dropping it down to 150 and trying again also had the warning

Logger: homeassistant.components.recorder.db_schema
Source: components/recorder/db_schema.py:581
Integration: Recorder (documentation, issues)
First occurred: 21:54:51 (2 occurrences)
Last logged: 22:07:15

State attributes for sensor.ai_automation_suggester_google_ai_automation_suggestions_google exceed maximum size of 16384 bytes. This can cause database performance issues; Attributes will not be stored

[pyrocumulus]

Same here, even with only 100 entities I get the same warning in the log. I tried getting more logs but even with the configuration.yaml change I do not see more data:

2026-06-23 23:29:20.431 DEBUG (MainThread) [custom_components.ai_automation_suggester.sensor] Sensor platform setup complete for provider: Google
2026-06-23 23:29:20.432 DEBUG (MainThread) [custom_components.ai_automation_suggester] Setup complete for AI Automation Suggester (Google) with provider Google
2026-06-23 23:31:25.621 DEBUG (MainThread) [custom_components.ai_automation_suggester.coordinator] Finished fetching ai_automation_suggester data in 4.732 seconds (success: True)
2026-06-23 23:31:25.623 WARNING (Recorder) [homeassistant.components.recorder.db_schema] State attributes for sensor.ai_automation_suggester_google_ai_automation_suggestions_google exceed maximum size of 16384 bytes. This can cause database performance issues; Attributes will not be stored

Obviously the debug level has been correctly set, I just don't see any raw request/response.

[JamesHannon]

I've switched to Anthropic (was trying to keep work and home separate), and that's been working without a hitch

However, be aware there's no free tier, but costs about $0.01 a hit to 4.6 sonnet.

[andynash]

Following this as I have the same problem.

Are the logs provided above sufficient to resolve this or is more information needed? Thanks for your work.

[ITSpecialist111]

Thanks all — @afreemo-glitch, @JamesHannon, @pyrocumulus, @andynash — for the extra data points. The recorder warning (State attributes for sensor.ai_automation_suggester_google_ai_automation_suggestions_google exceed maximum size of 16384 bytes) is a separate issue from the original "No automation YAML was returned" parser error, so I've addressed it directly.

Fix in v1.5.8 (d8fd997): the large suggestion payload attributes (suggestions, yaml_block, description, entities_processed, suggestion) are now marked as _unrecorded_attributes. They remain fully visible on the live state, in the Lovelace card, and through the HTTP API — they're just no longer written to the history database, which is where the 16 KiB cap lives. The warning should stop and no functionality changes.

For the original Google parse failure (@afreemo-glitch): I still need the raw provider response from debug logs to fix the underlying cause. @pyrocumulus — the reason you're not seeing request/response lines in debug is that the current build only logs the parser stage; I'll add raw response logging in the next pass so we can capture what Gemini is actually returning.

Please update to v1.5.8 via HACS and confirm the recorder warning is gone.

[andynash]

I've got this working thanks to Gemini.

Setting a custom system prompt is a great idea and definitely helps. You can use something like:

Return only valid YAML code. Do not add any explanatory text, introductions, conclusions, or Markdown formatting.

However, looking at the configuration, the root cause of the parsing errors was actually caused by the token limits:

Maximum input tokens (set to 500): This is extremely low for the Home Assistant context. Since the excluded domains/entities fields were empty, HA tried to grab absolutely all entities (sensors, lights, switches, etc.) and send them to Google as context. This doesn't fit into 500 tokens by a long shot. The prompt gets truncated halfway through, the AI never receives the complete instructions/schema, and it results in a broken or unparseable output.

    The Fix: Increase this limit to something like 32000 or 64000 (depending on the chosen model's capabilities). It is also highly recommended to populate the exclusion fields (e.g., excluding domains like weather, update, device_tracker) to keep the context size reasonable.

Maximum output tokens (set to 500): This might be enough for a very simple automation, but a slightly more complex YAML code (especially with multiple conditions) will get cut off before the generation finishes.

    The Fix: Raise this limit to at least 2048 or 4096.

After adjusting these token limits and providing the custom system prompt, the AI Automation Suggester works flawlessly!

Have you had a chance to check what this leads to in terms of API consumption/cost?

I guess it is not a huge deal if it is a reasonable cost as you only need to run this occasionally but it would be reassuring to be able to quantify it...

Also, not understanding fully how the integration works yet, if you limited it to say 50 automations from the automations.yaml, would it check the same first 50 each time, or use a different approach?

[Pan-Lorenc]

Hi! To answer your questions:

1. API Cost / Consumption
   I am currently using the free tier of the Google Gemini API (via Google AI Studio). Models like gemini-1.5-pro or gemini-1.5-flash have very generous free limits that are more than enough for generating automations occasionally. My current cost is exactly $0. Even if you were on a pay-as-you-go plan, sending a prompt with a large context window (e.g., 32,000 input tokens) costs just a fraction of a cent per request.

2. Limiting the scope
   I haven't tested the exact logic of limiting the add-on to 50 automations from automations.yaml. However, from my recent troubleshooting, I found that the most crucial part for saving tokens (and avoiding parsing errors) is limiting the entities exposed to the model, rather than the automations.

I highly recommend using the exclusion fields in the integration settings to filter out domains you don't need for automation generation (like weather, device_tracker, update, etc.). By doing this and manually including the specific entity_id in your text prompt when asking for a new automation, it works perfectly and keeps the token usage well within reasonable limits.

[Pan-Lorenc]

Here is the definitive proof of the bug in the "formatting repair" logic:

I captured the raw string that the parser tried to process after removing the markdown tags. It looks like this:

- id: '1780905558751'alias: pomocnik klima zapdescription: ''mode: singletrigger:- platform: stateentity_id: input_boolean.klimatizaceto: 'on'action:- action: button.presstarget:entity_id: button.air1_power_on- id: '1780905604476'alias: "Pomocn\xEDk klima vyp"description: ''mode: singletrigger:- platform: stateentity_id: input_boolean.klimatizaceto: 'off'action:- action: button.presstarget:entity_id: button.kancelar_air_power_off

Conclusion:
Gemini generates the correct YAML logic, but the add-on's internal regex/formatting repair aggressively strips out all newline characters (\n) and indentations along with the markdown tags. This flattens the multi-line YAML into a single unparseable string, which immediately causes the while parsing a block mapping ... expected <block end>, but found '<scalar>' error in Home Assistant.

[Pan-Lorenc]

Found a workaround via prompt!

If anyone is struggling with the YAML parsing errors (expected , but found '') when using Google Gemini, there is a very simple trick. Put this instruction into the Permanent custom system prompt in the add-on configuration:

Structure the YAML code just like AI Claude from Anthropic.

For some reason, instructing Gemini to mimic Claude's formatting style completely prevents it from using the destructive Markdown wrappers that break the integration's parser. The generated YAML is imported cleanly, with all newlines and indentation perfectly intact.