Home, Install, Choosing a Model, Basic Use, Bookmarklets, Prompt Reference, Prompt Formatting, API Switching, Chaining Inference
Clipboard Conqueror specific invoke and operator settings are defined in appSettings
- note: some links in these files are permanent links the code locations in the github repository, but because of their nature they link to old versions. Don't pay too much attention to line numbers, but they should help orient you when you are looking at the up to date source code on your computer.
Most of these settings can not overlap. If they do, Clipboard Conqueror may not function properly.
Define new system prompts in idents. These should be strings and use unique key names. (key : "text")
|||terminator:save|assistant takes on the role, personality and mannerisms of the Terminator as played by Danny Devito.
is equivalent to adding: terminator: "assistant takes on the role, personality and mannerisms of the Terminator as played by Danny Devito." to setup.js. though :save is less permanent and does not persist after restarting Clipboard Conqueror.
|||terminator:file| will write the prompt to disk. Once the 0prompts.json file exists, it overrides changes to idientities in setup.js
use it like: |||terminator| Will you be back?
The default system prompt is defined under persona, and |||anyValidPrompt, !anyname| both prevent sending the set default persona this query.
Don't be too intimidated when changing the code or json files, all the settings are simple objects. {key: "string value", key2: "this text is another string"}. There are numbers in the params, which are even simpler. {number: 123}
In the context of programming, text is usually stored as strings defined by "" quotes, '' apostrophes, or `` grave symbols. Its only intimidating if you're not used to it. Worst case you can download the zip again and start over. You can also try asking me if you're unsure about anything.
Prompts defined as objects must folow this format:
_nameYourPrompt: {
- the key name doesn't have to start with an underscore, but it helps keep them distinct from standard prompts. You do you.
one: { define the first turn of a complex prompt. You can have multiple turns that run in sequence. See two below.
- Top level object keys don't matter and can be named anything. They are shown in the console though.
Below are the valid key names for setting up a complex prompt. All keys are optional, an empty object will allow the first turn to run as normal. This is useful to keep things simple when setting up a response review prompt.
-
systemRole: "", equivalent to |||}name| These change the respective role names.
-
assistantRole:"", |||!name|
-
userRole: "", |||>name|
-
historyuserRole: "", Changes the previous turn's username, see the "]" operator
-
inferenceClient: "", Must match a key in endpoints.endpoints
-
format:"", Must be a valid format key in promptFormats
-
params:"", Must match a key in apiParams. It it is the same as the endpoint, then switching endpoints will change the parameters as well
-
jsonLevel: "", markup,full,keys,none. Default: none. Completion and chat combined endpoints only. full sends JSON.Stringify(allSetPrompts) into the system prompt. keys sends the contents like key : content \n\n key2 : ... markup makes each prompt it's own chat message with the key name as the role, none sends only the prompt text.
-
existingPrompts: ["use", "valid", "prompt"], names. To send a new custom prompt, you must first create a standard text string prompt. Send as many as you like. Equivalent to |||use,valid,prompt|
-
continue: "", define the start of the assistant response, uses the ~ operator
}, close the object, don't forget a comma.
two:{ The second turn has the same valid keys as one. The sequence continues to n as needed.
- existingPrompts: ["second", "Turn", "Prompts"],
},
n:{} add as many inference turns as you like.
}, don't forget to close the encompassimg object. Congratulations, you'v'e set up a new complex or multistage prompt
appSettings = {
- all these keys are required to be present.
invoke: "|||", could be anything # or AI: or Help Me ObiWan Kenobi, whatever you want.
endTag: "|", its the limiter after |||: promptf "|"quick prompt"|"query. This should not be the same as any below. Again, pretty much anything goes for non single charachter operators below.
save: "save", like |||name:save|
true: "true", like |||setting:true|, to ease use in other languages
false: "false", like |||setting:false|
saveAgentToFile: "file", like |||prompt:file| This will write 0prompts.json if it doesn't already exist.
delete:"delete", like |||prompt:delete|, removes the saved prompt from memory, //ToDo: Remove prompt from file.
settinglimit: ":", like |||prompt:save|
continueTag: "~~~", like ||| prompts| text ~~~ This text will start the assistant's response, and can support commas and colons without changing these settings.
batchContinueTag: "~", |||~ This text will start assistant's response, prompts| can support periods but not commas or colons, change agentSplit to enable commas. Must be a single charachter.
systemTag: "}", like |||} system Name| Must be a single charachter.
userTag: ">", like |||> User Name| Must be a single charachter.
assistantTag: "!", like |||! Assitant Name| Must be a single charachter.
backendSwitch: "$", |||$| is similar to |||ooba| with default config order, but only changes the endpoint, depending on endpoint settings. Must be a single charachter.
batchNameSwitch: "]", changes chat history user name next turn Must be a single charachter.
batchAssistantSwitch: ";", changes chat history assistant name Must be a single charachter. This one only works when chatting, not when chaining inference. use > to set the incoming text identity when chaining.
historyName : "continue",
batchSwitch: "@", like |||@name| Initiates chain of inference, sending the response as user for an additional turn. Must be a single charachter.
batchMiss: "#", like |||#@name| Must be a single charachter.
formatSwitch: "%", like |||%alpaca| changes only the prompt format. Do this one first before !>} Must be a single charachter.
paramSwitch: "^", like |||^ooba| changes only the parameters. Must be a single charachter.
batchLimiter: "", if empty, will mark the continue history with full format chat turns. Anything other than "" nothing will use this to seperate chat turns in the history.
setJsonLevel: "`", like |||`1| or |||`json| etc
empty: "empty", |||empty| to prevent sending the default system prompt.
emptyquick: "e", |||e| for a quicker empty system prompt.
agentSplit: ",", like |||prompt,write|
rootname: "system", when using ||||quick prompts| this key is used for json and partial modes. like |||`json| this text goes in the value| "request". example json system prompt: {"system": "this text goes in the value"}
paramatron: true, false disallows naked key name matching like |||ooba|
removeformatting: true, removes formating <|any|> from the response before sending to the clipboard. Eases using chatML ar llama3 formats with alpaca models.
clean: true, clean takes out the rootname key when it's not set. Set false to always send the rootname even if there is no associated text to send.
setInstruction: "PROMPT", like |||PROMPT:assistantRole| James Bond,
options: This is a comment to ease of configuring the .json file if written, can be ommitted.
setPromptFormat: "FORMAT", like |||FORMAT|chatML. If used like |||FORMAT:save| upi can copy correct json to set multiple prompt format segments at once. Not reccomended.
setParams: "PARAMS", like |||PARAMS|kobold
writeSave: "|||name:save|", writes this out when you do |||prompt,write|, purely textual for user convenience.
writeSettings: "|||FORMAT|", this is like writesave, it's to ease handling of written and pasted prompt formats.
writeSplit: "\n _______\n", similar to writeSave and writeSettings, adds text after |||name,write| idk, it felt neccessary. make it "" and its like it isnt there at all.
- all these keys are required to be present.
escapeReadkey: true, true enables the keyboard listener to cancel generation on escape press.
notifications: true, true enables notifications, no support for sound on linux.
writeFiles: false, true to write 0formats.json, 0prompts.json etc. |||name:file| will write 0prompts.json and it will then be used while present even if this is set false.
duplicateCheck: false, some other clipboard applications duplicate copied text back to the clipboard, set this true to catch those and preserve the proper last copied text. Clipboard Conqueror should work seamlessly alongside any other program with the right settings.
defaultClient: "kobold", must match a key in endpoints. Recommend using kobold or ooba. Completion endpoints support all features of CC.
defaultOptions: This is a comment to aid editing inside the json files when writeFiles is set true. It is never referenced in the software.
instructFormat: "llama3", overrides the defaultClient's set format. Must be a valid format key in promptFormats
instructOptions: This is a comment to aid editing inside the json files when writeFiles is set true. It is never referenced in the software.
persona: "default", must be a valid identity key in idents set as "e" for empty or no default system prompt. "pro" is a cleaner more professional prompt that thinks step by step.
- some of these keys are optional or unused depending on the type of endpoint. The prescence of some keys will change the request message strucure to match various api requirements.
Move your favorite endpoints to the top and invoke them by $ from top to $$$... at bottom.
Set endpoints persist until changed and can overwrite generation parameters
kobold:{ the object key is used to set the endpoint like |||kobold|
type: "completion", completion or chat, completion allows CC to control the formatting completely, chat sends openAI style messages, and a jinja template in a key drfined under templateStringKey
jsonSystem : "none", markup,full,keys,none. Completion and chat combined only //full sends JSON.Stringify(allSetPrompts) into the system prompt. keys sends the contents like key : content \n\n key2 : ... markup makes each prompt it's own chat message, none sends only the prompt text.
textHandle: "inputs", Replaces the "prompt" key with this value. Completion only till I spot a chat endpoint that doesn't use "messages" or someone asks for it.
paramPath: "options", sets up the generation parameters key for APIs that nest the parameters down a level like Ollama and Lorax
maxReturnTokensKey: "max_new_tokens", Set this to enable quick request length like |||500| for endpoints that use different names for the requested tokens like Ollama. This is required for endpoints that use paramPath.
buildType: unused for completion, combined should be used for chat, the others are experimental. options combined, system, or key
url : "http://127.0.0.1:5001/api/v1/generate/", then endpoint url.
params: "kobold", sets the generation parameters. must match a key in apiParams. It it is the same as the endpoint, then switching endpoints will change the parameters as well
autoConfigParams: false, false prevents overriding params with |||tgwchat| omitted = true, used to suppress param changes when switching endpoints
paramPath: "string", changes how the parameters are nested to the query. Ollama puts all parameters under "options" and Lorax puts them under "parameters" for example
maxReturnTokensKey: "num_predict", When present the string in this key is used to define the key used in the request for the max return tokens. |||500| If omitted the key is looked for in the params object from common request formats.
templateStringKey: "instruction_template_str", if present and not an empty string (""), sends a jinja template under this defined key name in the request. If ommited no jinja is sent.
format: "llama3", must be a valid instruction format from promptFormats.
autoConfigFormat: false, false prevents overriding prompt formatting with |||kobold| ommited = true, used to suppress prompt format changes when switching endpoints
textHandle: "text". changes the key used to send the prompt. Completion via replicate reqires text: {{prompt}} rather than prompt: {{prompt}}. This is used when the endpoint does not use the default prompt key "prompt". I can't remember if it can also change the "messages" key in chat endpoints.//ToDo: figure this out
systemLocation: "any/unused", Currently the prescence of this key will send the system promps at the top level of the request instead of as messages in the chat endpoint. Required for anthropic api.
noFormat: true, prevents sending the jinja template or kobold adapter, this breaks many features that change the prompt format such as !>}% operators. omitted = false. used to suppress param changes when switching endpoints and comply with strict request formatting.
model: optional model name sent with the params, overwrites model field if set in apiParams.
authHeader: "Authorization", optional authentication header for web api access.
authHeaderSecondary: "Bearer " optional secondary authentication header for web api access.,
key: "optional API authentication key for web api access. put your openAI key here and send to the openAI url to use ChatGPT apis",
Together these two authHeader options work together with the API key to create a single authorization header (Header{Authorization: Bearer Key})
headers [["key","value"],["key","value"]], optional headers for web api access, an array of key value pairs for custom headers. If you want, define your key in here. Example: [["X-Custom-Header", "Custom-Value"], ["Accept", "application/json"]].
error: "message", defines the response error text key to help verbosity of api errors.
outpoint defines the structure for recieving text back from the infrerence engine payload.
results[0].text: "results" goes in one, [two sends a number (0)], three is "text".
- outpointPathSteps: 3, key for a nifty switch case to get the response payload without modifying inference.js.
keys must be lowercase numbers up to ten like one two three four...
-
one: "results",
-
two: 0,
-
three: "text"
- outpointPathSteps: 4,
- one: "choices",
- two: 0,
- three: "message",
- four: "content"
- all of these keys are optional and empty "" if not defined.
completion has all the possible keys and needs none because they are blank, empty strings: "".
completion: { again the object key can be used like |||chatML| or |||%chatML|
order: ["system","user","assistant"], Completion endpoints only, controls turn order, only supports system, user, assistant, duplication is possible. If order is not defined, the default order, system, user, assistant, is applied.
only add these two (bos,eos) if you're sure you know what you're doing, duplicate bos may reduce response quality
bos: "", always goes first
eos: "",
they are generally added by the tokenizer. EOS goes on the end after responseStart, this may mess you up. Only use if neccesary to close BEFORE the generated assistant response.
The following to all turns, some formats have different turn initialization tokens so startTurn, endTurn, and endRole may not be appropriate for all formats. This is to avoid repeating the same info over and over again.
startTurn: "", use if startSystem, startUser, and startAssistant are all the same
endTurn: "", use if endSystemTurn, endUserTurn, and endAssistantTurn are the same.
endRole: "", use if all roles use the same role name closure.
roleGap: "", For adding a gap between role identifier and content. For individual gaps put them on endSystemRole, endUserRole, endAssistantRole.
below, the above keys are commented in the places they append. The following is in the order that each is appended in a turn with the default system, user, assistant turn order.
startTurn
startSystem: "",
systemRole: "",
endSystemRole: "",
endRole
rolegap
prependPrompt: "",
systemAfterPrepend: "",
system prompt text inserted here, chat history is added at the end of the system prompt text.
postPrompt: "",
memorySystem: "",
endSystemTurn: "",
startTurn
startUser: "",
userRole: "",
endUserRole: "",
endRole
rolegap
memoryUser: "",
user prompt inserted here
endUserTurn: "",
endTurn
startTurn
startAssistant: "",
assistantRole: "",
endAssistantRole: "",
endRole
roleGap
responseStart: "", This segment isn't correctly being used by passed jinja templates.
endAssistantTurn: "", only used for history markup
endTurn
specialInstructions: "" only for jinja template
- note: order of keys doesn't matter. It will get where it belongs if you use the right keys.
- startTurn: "<|start_header_id|>",
- endTurn: "<|eot_id|>",
- endRole: "<|end_header_id|>",
- roleGap: "\n\n",
- systemRole: "system",
- userRole: "user",
- assistantRole: "assistant",
- startTurn: "<|im_start|>",
- endTurn: "<|im_end|>\n",
- systemRole: "system",
- userRole: "user",
- assistantRole: "assistant",
- roleGap: "\n",
- endTurn: "\n\n",
- systemRole: "Below is an instruction that describes a task. Write a response that - appropriately completes the request.",
- userRole: "### Instruction:",
- assistantRole: "### Response:",
Different inference apis may have different key names and requirements. Notably Kobold uses max_length while TextGenWebUi uses max_tokens to define the maximum response length. You may need to reference the docs for your chosen inference server to determine the correct parameter keys. The required keys may also change between completion and openai chat endpoints on the same inference server.
Clipboard Conqueror should work with any inference API, or as many parameter sets as you care to define for quick swapping.
If you find an api that doesnt work, or need help, contact me at [email protected] or open an issue or discussion on github.
ooba: { again, the object key is used like |||ooba|
-
max_context_length: 8192, may not be required, CC relies on the inference API to trim overlong context. Kobold intelligently preserves the system prompt.
-
max_tokens: 700, Number of tokens requested. This number is not strict, it's the maximum. It may be less if the model stops generating.
-
temperature: 1, Temperature changes scaling of final token probability, less than one makes unlikely tokens less likely, more than one makes unlikely tokens more likely, normalizing final probabilities.
-
repetition_penalty: 1.05, rep penalty helps prevent repetition. 1.00 is no penalty. High values can negatively impact formatting charachters.
-
top_p: 1, discard possible tokens by throwing out lest likely answers. 0.8 throws away least likely 20%
-
top_k: 0, discard all but top_k possible tokens. top_k: 3 means all but the most probable 3 possible tokens are discarded.
-
top_a: 0,If the maximum probability is very high, fewer tokens will be kept. If the maximum probability is very close to the other probabilities, more tokens will be kept.
-
typical: 1, this allows top tokens to be suppressed in favor of slightly lower probability tokens. May improve quality and control repetition.
-
tfs: 0.97, tail free sampling, removes unlikely tokens from possibilities by finding the platau where tokens approach equally unlikely. 0.99 maximum. Higher value finds a lower, flatter plateau.
-
min_p: 0.1, discard possible tokens less than 10% as likely as the most likely possible token. If top token is 10% likely, tokens less than 1% are discarded. I typically only use min_p.
-
repetition_penalty_range: 512, how many tokens to check for repetition penalty, longer slows generation.
-
repetition_penalty_slope: 0.7, this applies a curve to repetition penalty so that tokens farther back in the context are penalized less.
-
sampler_order: [6, 5, 0, 1, 3, 4, 2] the sampler order. Only adjust this if you know what you're doing.
}
Home, Install, Choosing a Model, Basic Use, Bookmarklets, Prompt Reference, Prompt Formatting, API Switching, Chaining Inference