AI analysis returned no valid data

[Sorry-Im-Stupid]

Almost every message I try to have it make an entry for gives me that.
Is there a minimum message length it needs or something?
Is there no way to have it analyze multiple messages at once?
The automatic functions aren't doing anything either.
I've also noticed the defaults for everything assume you're connected to an API with 10k plus token context, but there's no setting for running local, which for the model I use is limited to 8k token context.
The default for the auxiliary also assume an API, and when I set it to fallback it puts up an error message anyway.

[SenriYuki]

Hi, thanks for the report. A few clarifications:

About "AI analysis returned no valid data": This message is a bit misleading. It does not mean the AI returned nothing — it means the AI did return content, but the content did not contain any of the required structured tags (... / ... / / ). The plugin parses those tags to fill the timeline/NPC/items/etc., so if the model replies in free-form prose without those wrappers, nothing can be extracted.

Why this happens on local 8k models: The analysis prompt asks the model to output a fairly strict multi-field tag block (time / location / atmosphere / characters / costume / item / affection / npc / agenda + a separate block). Small local models with ~8k context generally do not have strong enough instruction-following to reliably produce that structure, so they default to writing natural-language summaries — which then fail the tag check. This is a model-capability limitation, not a plugin bug.

Note on Gemini-specific fix: v1.14.1 added a Gemini native-API path because Gemini via OpenAI-compatible relays was dropping safetySettings / systemInstruction and frequently returning prose instead of tags. If you happen to be using Gemini through a relay, please update to v1.14.1+. For local 8k models the underlying cause is different (model capability), and the fix above won't help.

Your other questions:

Minimum message length: No, there is no minimum-length gate. The check only looks for the tag wrappers.
Analyzing multiple messages at once: The wand button is single-message analysis by design. For bulk processing, use "Batch Summary" (scan range) and "Auto-Fill Timeline" — those are the multi-message paths.
Automatic functions doing nothing: Same root cause. If the model can't produce tags on a manual run, the auto-fill / on-send completion paths will also silently produce nothing.
Defaults assume large context: Correct that the built-in prompts target larger models. For 8k local models I'd suggest shortening the analysis prompt in Settings → Prompts → AI Analysis Prompt (e.g. keep only time / location / characters / event and drop costume/agenda/npc/affection), which both fits 8k and is much easier for small models to follow.
Aux API "error" when fallback is on: That toast is a yellow warning, not an error — it just notifies you that the aux API failed and the request has already been re-routed to the main API. The request itself does complete via the main API in that case.
Recommended setup if you must stay on a local 8k model:

Set a stronger remote model as the Auxiliary API (used only for analysis/summary), and keep the local model as the main API for roleplay. Auxiliary API → "Use for AI analysis / wand / pre-send timeline fill" is on by default.
Or significantly trim the AI Analysis Prompt to a minimal field set as described above.
Leave "Fallback to main API on aux failure" off unless you're sure you want the local model to be hit as backup.

[Asherathe]

I use Horae with local models, and it's very usable with a model that's good at instruction-following and formatting, and doesn't over-think.

Models that work well, IME: Gemma 4 31B, Qwen 3 32B or 3.5 27B, GLM 4 32B

An IQ3 quant allows 20k context on my very modest system, and is still smart enough to write the tags properly. You might have to experiment with models a little bit. If you use RP-specific finetunes and merges, one might not work well and another very similar one might work fine.

[Sorry-Im-Stupid]

I'm not using an off-site API, I don't have the money for that.
I found if I cranked up the response length of the messages it was analyzing it did more often generate useful data, so that is a factor in its' efficacy.

[SenriYuki]

Hi @Asherathe, thanks so much for sharing this! It’s genuinely useful field data. Since I mostly work with hosted models day-to-day due to my preference for long context, I haven't benchmarked Horae's tag extraction systematically on local setups.

Your "model + quant tier + context length" rundown is incredibly valuable — not just for me, but also as a great reference for any local-model users who run into this issue in the future.

[SenriYuki]

Your observation is genuinely useful — "if I crank up the response length on the analyzed message, extraction success goes up" lines up perfectly with what @Asherathe pointed out above. When the message is short, the small model has to immediately switch into tag mode and fill every required field in a single shot, which is structurally hard for it. When the output length is higher, the model has more breathing room, making writing those extra structured tags a natural continuation rather than a harsh context switch.

Since paid APIs aren't an option, here's a combo worth trying — all fully local:

1. Switch to a model that's good at structured output: Use @Asherathe's list as a starting point (Gemma 31B / Qwen 3 32B / Qwen 3.5 27B / GLM 4 32B). These are generally regarded as strong on instruction-following and less prone to "over-thinking." If VRAM is tight, drop to IQ3 / IQ4 quants — @Asherathe reported hitting 20k context even on fairly modest hardware that way.

2. Trim the prompt: Go to Settings → Prompts → AI Analysis Prompt — delete the optional fields (costume / agenda / npc / affection) entirely and keep only the essentials (time / location / characters / event). Fewer fields = a tighter template = much easier for a small model to follow.

3. Push the max response tokens higher: You already saw this work. Give the model enough headroom to actually finish writing the tags without getting cut off.

4. Use "Batch Summary" instead of the single - message wand: Give it a range (say, the last 10 messages). With more context per pass, the model has a much easier time consistently identifying what needs to be extracted.

If none of the above gets the tags working, feel free to paste the raw AI output from a failed attempt (the actual text the AI produced before the extraction failed). That will help me tell at a glance whether the model didn't enter tag mode at all, or if it did write the tags but with a slight formatting quirk that the parser missed. The fix is different for each case.

[Asherathe]

The only things local models are consistently bad at are detecting when food or drink have been consumed and removing them from the item tracker, and detecting when agenda entries have been completed. I always have to do these manually. Detecting mood, clothing, and atmosphere is occasionally unpredictable, but has improved a lot with recent updates; some models are better at these than others. Updating date, time, and location, and generating event summaries works very well with every model I've used.

[SenriYuki]

Hi @Asherathe — same experience here, and notably this isn't a local-only ceiling. I run Horae mostly on hosted frontier models and the exact two failure modes you listed are still the most stubborn ones even there:

• Consumable items not being decremented/removed —drinks and small misc. items in particular. My personal favorite "gallows humor" case so far: a character spilled an urn of ashes (bottle visibly empty in the scene) that it didn't removed by AI, and after the next shower scene the bottle was magically full again, with water.
• Agenda entries not closing out even when the goal is clearly fulfilled in-scene.

Pretty much everything else (time / location / event summaries / costume / atmosphere / NPC tracking) is solid on a strong remote model.

My take: the prompts themselves are already in a reasonable spot — these two edge cases are essentially asking the model to do subtractive/negative state tracking ("this no longer exists" or "this is no longer pending"). LLMs are notoriously worse at processing non-existence or deletion than they are at additive tracking (adding new items/NPCs). I'd expect this to quietly improve as base models get smarter, rather than needing more prompt surgery.

A couple of ways I handle this in the meantime:

• A tiny custom nudge: For agendas, sometimes adding a specific phrase to the prompt or a system note reminding the model to "explicitly check for completed objectives" helps, though it's still hit-or-miss.
• Quick manual overrides: For inventory and agenda misses, since the model already generated the bulk of the data correctly via the automation/wand, doing a quick 2-second manual click to delete the item or check the agenda box in the UI remains the most reliable fix for now.