Files & Images
Telosnex lets you share files and photos with AI.
You can drag and drop files onto Telosnex to add them. Or, tap the Files button.
Telosnex does not store your files or upload them to the cloud. They are only sent to the AI, and Telosnex only integrates with AI providers that respect your privacy.
Files Menu in Companion
- View: Press a file or photo to view it.
- Remove: Press the remove icon to remove a file or photo.
- Read
- AI Select: Use the prompt to pick relevant information from files. If this is chosen, you can also choose how many pages to select.
- Full Scan: Send the entire file as the prompt. If this is chosen:
- An estimate of cost & time, based on the current Script, is shown. If Auto is selected, the estimate is based on your AI provider.
- You can also choose how files are split. (see Files: Full Scan)
- Photos:
- The photo is sent to AI, with the prompt, and the AI is asked to analyze the photo. Then, the AI is asked to answer the prompt.
- If your AI provider supports photos, the photos are sent to that AI. If not, the photos are sent to Anthropic's Claude Haiku, the cheapest AI that supports photos.
tip
On Android & iOS devices with cameras, there is also a Take Photo button.
Files: AI Select
When you attach files and choose AI Select, an AI running on your device is used to select relevant information from the file. Also, Telosnex lets you select how much relevant information to send.
These are unique features of Telosnex.
Virtually all AI providers process files and websites by:
- Uploading your files to their servers.
- Run an AI on the server that turns the files into embeddings, a numerical representation of the text.
- Compare the file embeddings to your prompt's embedding, rank them, and pick the most relevant.
- Send your prompt and the most relevant text to the LLM. (chat AI)
This causes significant conflicts of interest that compromise quality, control, and privacy:
- It is cheaper to only select a few relevant snippets to send to the chat AI.
- It is cheaper and faster to store the file chunks and their embeddings on the server, so they can be reused for future prompts, instead of requiring their server to reprocess the file every time.
Telosnex's core principles include Transparency, Control, and Privacy. When developing, we must always choose to create technological solutions to avoid creating inherent conflict of interest.
Therefore, we invested deeply to find a way that runs the embeddings AI right on your device. On every platform. We do not store them or your files, even on your device.
Files: Full Scan
When you attach files and choose Full Scan, the file is split into multiple chunks, and each is sent with your prompt to AI.
Controlling the # of pages in the chunks helps with:
- Large tasks that have as much output as input: For example, rewriting a book. Most AI providers are limited to responding with 12 pages.
- Speed: It is faster to have the AI look at 1 page at a time 10 times and answer a question on each page, than have it answer 10 questions at once.
- Helping AI focus: AI's memory has increased dramatically. For example, OpenAI's best in 2022 was 13 pages. Now, it is 384 pages. However, AI is still trained to answer concisely, and much like us, can miss details when going through large amounts of text. Working on smaller pieces makes it more likely to be accurate.
Supported File Types
You can attach these file types:
- Photos: .png, .jpeg, .jpg, .psd, .webp, .bmp, .gif, .ico, .tiff, .apng, .tga, .cur
- Spreadsheets: .csv, .tsv, and Microsoft Excel (.xlsx, .xls)
- Documents: Microsoft Word (.docx, .doc), PDF (.pdf), Text (.txt), HTML (.html), CSS (.css), SVG (.svg)
- URLs: Telosnex will download the page and convert it to text.
- Other/Code: Telosnex accepts any files that are text.
- URLs (web pages, Telosnex will download the page and convert it to text)
Examples
- Find information in the file: It is unlikely the AI will respond with 13 pages of information. So, we don't need to split the file into smaller pieces.
- Rewrite the file: For example, rewriting Moby Dick using Generation Z vocabulary. Ideally, the AI response has as many words as the original book, and when testing, we noticed it tended to skip some parts. Decreasing the # of pages in each split makes AI much more likely to respond at length.
Telosnex has a maximum limit of 12 pages per split.
That is because most AI providers limit responses to 12 pages.
Even provider who don't get bad results: AI is trained to answer concisely. Testing over months has shown it is impossible to get extremely long response.
Setting the default to 12 pages ensures that you don't have to remember all of this, and fiddle around with controls to get good answers.
This also saves significant amounts of money.
If you sent a 100 page file to GPT-4, and the answer was too short, and then you set a split, you would have wasted about a $1 on the bad answer.