Chats

Chats allow you to ask anything with your sources as context. When all sources are selected, Decosmic searches across the space and keeps the retrieved context balanced across matching files instead of letting one file crowd out every other result.

If you would like to chat with specific sources, open the source list and select one or more files. The request is then restricted to those files, which is useful when comparing language versions or working with a small group of related documents. Multi-source requests retrieve a deeper set of up to 15 balanced source rows so relevant entries from each selected file can reach the model. Each selected file gets its own retrieval pass, with a phrase-matching fallback for exact wording when a file's vector index is incomplete. Single-source requests remain capped at 5 rows. You can also select All Sources or Exclude Sources.

You can also select Exclude Sources if you decide to chat with the vanilla model with no context from your sources.

Agentic source search

When sources are enabled, Decosmic uses a bounded source-search agent. It seeds the first search from the current task (including a translation task referred to by a follow-up such as “use the tools”), then can make up to two additional focused searches when a phrase, language, or field is missing. If you ask for “all” languages, that means all language variants represented by the selected source files. The model receives the returned rows as evidence and can either quote a populated source value exactly or compose a new answer from relevant source terminology and precedents.

The loop is capped at four model steps (the initial search, up to two refinements, and the answer) and each search is read-only, limited to the files selected in the source picker, and protected by a timeout. Text returned by a source is treated as data rather than instructions. This lets the agent recover related rows without retrying forever or allowing a source row to change Decosmic's behavior.

Image Upload

You can upload a single image of not more than 5 MB and include it in a message to ask for anything related to the image.

Multiple image & file upload support will be coming later this year.

Web Search

By enabling the Web Search toggle, Decosmic will attempt to search the web for the most relevant search results relating to your message.

The references used will be cited in its response.

References

While a source-enabled response is running, the chat shows Searching sources... for the first retrieval and Searching deeper... for later refinement searches, then returns to Thinking... while the model writes the answer. After the message is finished, its references appear directly below it as one source chip per file. Select a file chip to open a dialog listing the matching source rows, scores, and any available web links. When a follow-up asks about the source, previously retrieved document entries are preserved as context so Decosmic can distinguish an exact quotation from a newly composed answer grounded in those entries. Web search results will also be included in this list if enabled.

Document references are stored with the chat message and are restored when the chat is reopened.

Models

Standard

openai/gpt-5.6-luna is the fastest, most cost-efficient model in the GPT-5.6 family, with vision capabilities for everyday work.

Super

openai/gpt-5.6-sol is the flagship multimodal model for complex reasoning, coding, knowledge work, and nuanced writing.

Custom

Select a custom model from the list of available models. These models are for advanced users who know what they're doing and want to experiment with different models.

Collaboration

Chats are currently not shared between team members.