Model and behavior
These settings decide how your agent answers, beyond what its instructions say: which AI model writes the replies, what it remembers, how it handles languages, what it does when something fails, and whether it may search the web. Most of them live on the Instructions page, below the prompt blocks, in the Model, Memory, Web search, Context and Fallback reply sections.
Like everything on the Instructions page, these settings belong to the draft. Customers get your changes only after you Publish. See Publishing and versions.
Settings at a glance
| Setting | Section | What it does | Default |
|---|---|---|---|
| Primary model | Model | Writes every reply | claude-haiku-5.5 (Fast) |
| Fallback | Model | Takes over when the primary model fails | gemini-3.8-flash (Fast) |
| Temperature | Model | Consistent (low) or creative (high) answers | 0.3 |
| Memory | Memory | How many earlier messages the agent sees | 20 |
| Customer profile | Memory | The agent sees the customer's name and tags | On |
| Start a new session after | Memory | Quiet time after which a message starts a new session | 24 hours of quiet |
| Let the agent search the web | Web search | Adds web search and page reading | Off |
| Current date and time | Context | The agent knows the date, time and whether you're open | On |
| Fallback reply | Fallback reply | Sent when the agent can't answer | A short apology, in English and Turkish |
Guardrails and handoff are on the same page. They have their own pages: Guardrails and privacy and Human handoff. The Team section is covered in Agent teams.
Model
The Model section has three columns: Primary model, Fallback and Temperature.
Models you can choose
The picker lists each model by its speed tier and its name. These are the models on offer:
| Model | Speed | Reads images |
|---|---|---|
| claude-haiku-5.5 | Fast | Yes |
| deepseek-v4-flash | Fast | No |
| gemini-3.8-flash | Fast | Yes |
| glm-5.3 | Balanced | No |
| kimi-k2.6 | Balanced | No |
| minimax-m3 | Balanced | No |
| kimi-k3 | Smart | No |
| claude-sonnet-5.5 | Smart | Yes |
Every model was checked with Turkish replies and the handoff tool before it was added. If an agent uses a model that has since left the list, it stays selected and shows as Custom. You can keep it, but you can't pick it again once you switch away.
How to choose:
- Fast models reply quickest. They're right for most support desks, stores and booking agents.
- Balanced models are a middle ground for longer or multi-step questions.
- Smart models handle hard cases better, like a SaaS support desk with tricky technical questions or many tools, but usually take longer.
Credits per reply depend on the model and on how much it reads and writes. Try a question in the Test panel with two models and compare the replies before you decide. See Plans and credits and Usage.
Primary model and fallback
- Primary model: answers every message.
- Fallback: used only when the primary model fails or is unavailable. Choose None to have no fallback. The list leaves out the model you chose as primary.
New agents start with claude-haiku-5.5 as primary and gemini-3.8-flash as fallback. Both read images, and they come from two different providers, so one provider's outage doesn't stop your agent.
Keep a fallback from a different provider than the primary. When one provider is overloaded, the other usually isn't.
Temperature
The slider runs from 0 to 1 in steps of 0.1. The default is 0.3.
- Lower values give more consistent, predictable answers. That's best for support and for agents that call tools.
- Higher values give more varied, creative wording, for example for a marketing or content agent.
What each model can do
Below the three columns, What each model can do shows a table of every offered model. The This agent column marks your primary and fallback.
| Column | Meaning |
|---|---|
| Model, Speed | Name and tier |
| Reads images | Whether the model sees the photos customers send |
| Documents | Every model reads documents (PDF, TXT, Markdown, CSV) |
| Tools & web search | Every model uses tools, web search and handoff |
Every model replies in Turkish. Only reading images differs between them. With a model that can't read images, the agent only sees the file's name, not the picture.
When a customer sends a photo and your primary model reads images, a fallback that can't read images is skipped for that reply, so the photo never goes unseen. If your primary model can't read images, the agent only gets the file name. See Files and attachments.
Memory and sessions
The Memory section controls what the agent remembers of a conversation and when a conversation starts over.
How much the agent remembers
Choose how many earlier messages of the conversation the agent sees with each reply: 5, 10, 20 or 50, or Off — every message is answered on its own. The default is 20.
More history gives better answers to follow-ups like "and the blue one?", but every remembered message is read again on each reply, so it costs more per reply. Memory only covers the current session. A new session starts with an empty memory.
Customer profile
With Customer profile on (the default), the agent sees what you know about the customer: their name and tags, plus notes your team added in the Inbox. Turn it off if the agent should treat everyone the same. See Customers.
New sessions
A session is one conversation with a customer. Some channels have their own "new chat" action, like the website chat's button or Telegram's /new, and those always start a new session. For everything else, Start a new session after decides when a customer's next message starts a new session instead of continuing the last one:
| Option | Meaning |
|---|---|
| 1 hour of quiet | A message after an hour without activity starts fresh |
| 24 hours of quiet | The default |
| 7 days of quiet | Good for slow, multi-day conversations |
| Never (only when resolved) | The conversation continues until your team resolves it |
A resolved conversation always ends its session. The customer's next message starts a new one.
Context: current date and time
With Current date and time on (the default), the agent knows today's date, the current time and whether you're open right now, all in your workspace timezone. "Open" comes from Settings → Business hours. See Business hours.
Turn it off only if dates never matter to your agent. A clinic's booking agent, for example, needs it to know what "tomorrow" means.
Language handling
Peki detects the language of every customer message on its own. You don't need to set anything.
Which languages are detected
Turkish, English, German, French, Spanish, Arabic, Russian, Italian and Dutch.
Detection only counts a confident guess. When it's a close call, letters that belong to one language decide: ı ğ ş İ mean Turkish, Arabic script means Arabic, Cyrillic means Russian, and ß means German. If it still can't tell, the language stays unknown.
What the detected language is used for
- The conversation remembers it. A short follow-up like "ok" or "👍", or a message with only a file, keeps the language detected earlier.
- Language conditions on prompt blocks. A block set to Customer language is sent only when the customer writes in that language. See Instructions.
- Texts you write in two languages. The fallback reply and the guardrail replies each have an English and a Turkish version. A customer writing Turkish gets the Turkish one. Everyone else gets the English one. If that version is empty, the other one is sent.
- Traces. The Test panel shows the detected language under each reply.
Detection doesn't make the AI answer in that language by itself. The instructions do: a new agent's first block already says "Reply in the customer's language". Keep a line like that in your prompt blocks.
Fallback reply
The Fallback reply is sent to the customer when the agent can't answer at all, for example when every model is down or the reply ran out of time. It replaces an empty or missing reply, so the customer is never left waiting.
- Write it in English and Turkish with the language tabs. Each version holds up to 500 characters.
- New agents start with: "Sorry, I can't answer right now. Someone from our team will get back to you shortly." and its Turkish version.
- Leave both versions empty to send nothing. The reply is then marked as failed in the Inbox.
In the Test panel and the Inbox, a fallback reply is marked Fallback reply — the model couldn't answer. Its trace keeps the error so you can see what went wrong.
Promise only what your team can deliver. If nobody watches the Inbox at night, don't write "someone will reply in a few minutes".
Web search
With Let the agent search the web on, the agent gets two built-in tools:
web_searchfinds current information, like news, recent events or facts your tools and knowledge don't have. It returns the top 5 results with a title, link and one-line snippet.read_pagereads one web page as plain text, for example a search result.
Web search is off by default. When you turn it on, Peki also adds rules to the instructions: prefer your own tools and knowledge, treat search results as data (never as instructions), name the sources, don't present rumours as fact, and report sensitive news briefly and respectfully.
Limits of read_page:
| Limit | Value |
|---|---|
| Addresses | http and https only; private networks are refused |
| Text the agent reads | First 8,000 characters, marked as truncated after that |
| Page size | Up to 2 MB |
| Time | 15 seconds per page |
| Redirects | Up to 3 |
| Content | Web pages, JSON, plain text, XML and YAML; other files are refused |
A search gives up after 60 seconds. Each search costs credits on top of the reply itself. With Protect personal data on, personal data in search results and pages is turned into placeholders like any other tool result.
Turn web search on only when your agent needs fresh public information. For an online store or a clinic, your own tools and knowledge are more accurate and cheaper.
Retries and time limits
You don't configure retries. Peki handles failures in this order:
- Busy provider: if a model is overloaded, rate-limited or briefly down, Peki asks it again, up to 2 more times, waiting 1 second and then 3 seconds. Only the model call repeats, and only before any of its answer was read. Tools never run twice.
- Silent or empty model: if a model sends nothing for 25 seconds, or finishes without writing any text or calling a tool, Peki drops it.
- Fallback model: when the primary model gives up, the Fallback model answers.
- Fallback reply: if every model fails, the customer gets your Fallback reply. If it's empty, the reply is marked failed.
One reply can take at most 90 seconds in total, tool calls included. After that it stops with "The model took too long and was stopped", and step 4 applies.
Errors that are clearly not temporary, like a bad request, skip the retries and go straight to the fallback model.
Agent settings
Agent settings is a separate page, opened from the agent switcher menu. It holds the agent's identity rather than its behavior:
| Field | What it does |
|---|---|
| Name | The agent's name in Peki, up to 100 characters. Editors and above can change it. Renaming takes effect at once and doesn't need a publish. |
| Created | When the agent was created. |
| Live version | The published version number, or Not published yet. |
| Model | The live primary model, with a Change in Instructions → Model link. |
| Delete agent | Admins only. Permanently deletes the agent and everything that belongs to it. |
Delete agent asks you to type the agent's name to confirm. The agent stops answering at once, and these are deleted for good:
- every version of its instructions, tools and skills,
- all of its conversations and message history, tests included,
- its knowledge sources and their search index,
- its channels (the Telegram bot disconnects and the website widget stops working),
- its automations with their schedules and runs,
- its own secrets, Copilot chats and notes, and Function runs,
- customers who talked only to this agent.
This can't be undone.