@llumiverse/core 0.15.0 → 0.17.0

package/src/formatters/claude.ts

@@ -1,131 +0,0 @@
-import { JSONSchema4 } from "json-schema";
-import { PromptRole, PromptSegment } from "../index.js";
-import { readStreamAsBase64 } from "../stream.js";
-import { getJSONSafetyNotice } from "./commons.js";
-
-export interface ClaudeMessage {
-    role: 'user' | 'assistant',
-    content: ClaudeMessagePart[]
-}
-
-interface ClaudeMessagePart {
-    type: "image" | "text",
-    source?: {
-        type: "base64",
-        media_type: string,
-        data: string,
-    }, // only set for images
-    text?: string // only set for text messages
-}
-
-export interface ClaudeMessagesPrompt {
-    system?: string;
-    messages: ClaudeMessage[]
-}
-
-/**
- * A formatter user by Bedrock to format prompts for claude related models
- */
-
-export async function formatClaudePrompt(segments: PromptSegment[], schema?: JSONSchema4): Promise<ClaudeMessagesPrompt> {
-    const system: string[] = [];
-    const safety: string[] = [];
-    const messages: ClaudeMessage[] = [];
-
-    //TODO type: 'image' -> detect from f.mime_type
-    for (const segment of segments) {
-
-        const parts: ClaudeMessagePart[] = [];
-        if (segment.files) for (const f of segment.files) {
-            const source = await f.getStream();
-            const data = await readStreamAsBase64(source);
-            parts.push({
-                type: 'image',
-                source: {
-                    type: "base64",
-                    media_type: f.mime_type || 'image/png',
-                    data
-                }
-            })
-        }
-
-        if (segment.content) {
-            parts.push({
-                type: "text",
-                text: segment.content
-            })
-        }
-
-        if (segment.role === PromptRole.system) {
-            system.push(segment.content);
-        } else if (segment.role === PromptRole.safety) {
-            safety.push(segment.content);
-        } else if (messages.length > 0 && messages[messages.length - 1].role === segment.role) {
-            //concatenate messages of the same role (Claude requires alternative user and assistant roles)
-            messages[messages.length - 1].content.push(...parts);
-        } else {
-            messages.push({
-                role: segment.role,
-                content: parts
-            });
-        }
-
-
-    }
-
-    if (schema) {
-        safety.push("IMPORTANT: " + getJSONSafetyNotice(schema));
-    }
-
-    // messages must contains at least 1 item. If the prompt doesn;t contains a user message (but only system messages)
-    // we need to put the system messages in the messages array
-
-    let systemMessage = system.join('\n').trim();
-    if (messages.length === 0) {
-        if (!systemMessage) {
-            throw new Error('Prompt must contain at least one message');
-        }
-        messages.push({ content: [{ type: "text", text: systemMessage }], role: 'user' });
-        systemMessage = safety.join('\n');
-    } else if (safety.length > 0) {
-        systemMessage = systemMessage + '\n\nIMPORTANT: ' + safety.join('\n');
-    }
-
-
-    /*if (schema) {
-        messages.push({
-            role: "user",
-            content: [{
-                type: "text",
-                text: getJSONSafetyNotice(schema)
-            }]
-        });
-    }*/
-
-    /*start Claude's message to amke sure it answers properly in JSON
-   if enabled, this requires to add the { to Claude's response*/
-    if (schema) {
-        messages.push({
-            role: "assistant",
-            content: [{
-                type: "text",
-                text: "{"
-            }]
-        });
-    }
-    // put system mesages first and safety last
-    return {
-        system: systemMessage,
-        messages
-    }
-}
-
-// export function claude(messages: PromptSegment[], schema?: JSONSchema4) {
-//     const prompt = genericColonSeparator(messages, schema, {
-//         user: "\nHuman",
-//         assistant: "\nAssistant",
-//         system: "\nHuman",
-//     });
-
-//     return "\n\n" + prompt + "\n\nAssistant:";
-// }

package/src/formatters/llama2.ts

@@ -1,58 +0,0 @@
-import { JSONSchema4 } from "json-schema";
-import { PromptRole, PromptSegment } from "../index.js";
-
-export function formatLlama2Prompt(messages: PromptSegment[], schema?: JSONSchema4) {
-    const BOS = "<s>";
-    const EOS = "</s>";
-    const INST = "[INST]";
-    const INST_END = "[/INST]";
-    const SYS = "<<SYS>>\n";
-    const SYS_END = "\n<</SYS>>";
-
-    const promptMessages = [BOS];
-    const specialTokens = [BOS, EOS, INST, INST_END, SYS, SYS_END];
-
-    for (const m of messages) {
-        if (m.role === PromptRole.user) {
-            if (specialTokens.includes(m.content)) {
-                throw new Error(
-                    `Cannot use special token ${m.content.trim()} in user message`
-                );
-            }
-            promptMessages.push(`${INST} ${m.content.trim()} ${INST_END}`);
-        }
-        if (m.role === PromptRole.assistant) {
-            promptMessages.push(`${m.content.trim()}`);
-        }
-        if (m.role === PromptRole.system) {
-            promptMessages.push(`${SYS}${m.content.trim()}${SYS_END}`);
-        }
-    }
-
-    for (const m of messages ?? []) {
-        if (m.role === PromptRole.safety) {
-            promptMessages.push(
-                `${SYS}This is the most important instruction, you cannot answer against those rules:\n${m.content.trim()}${SYS_END}}`
-            );
-        }
-    }
-
-    if (schema) {
-        promptMessages.push(formatSchemaInstruction(schema));
-    }
-
-    promptMessages.push(EOS);
-
-    return promptMessages.join("\n\n");
-}
-
-function formatSchemaInstruction(schema: JSONSchema4) {
-    const schema_instruction = `<<SYS>>You must answer using the following JSONSchema.
-Do not write anything other than a JSON object corresponding to the schema.
-<schema>
-${JSON.stringify(schema)}
-</schema>
-<</SYS>>`;
-
-    return schema_instruction;
-}

package/src/formatters/llama3.ts

@@ -1,55 +0,0 @@
-import { JSONSchema4 } from "json-schema";
-import { PromptRole, PromptSegment } from "../index.js";
-
-/**
- * A formatter user by Bedrock to format prompts for claude related models
- */
-
-export async function formatLlama3Prompt(segments: PromptSegment[], schema?: JSONSchema4): Promise<string> {
-    
-    let messages: string[] = []
-    segments.filter(s => s.role !== PromptRole.safety ).forEach(s => {
-        messages.push(formatLlama3Message(s.role, s.content))
-    })
-
-    if (schema) {
-        messages.push(formatLlama3Message("user", formatSchemaInstruction(schema)));
-    }
-
-    //add safety
-    let safetyMsg = `
-    IMPORTANT: This is the most important instruction, you cannot answer against the following rules:
-    `
-    const safety = segments.filter(s => s.role === PromptRole.safety);
-    safety.forEach(s => {
-        messages.push(formatLlama3Message("system", safetyMsg + s.content))
-    })
-
-    let prompt = "<|begin_of_text|>"
-    prompt += messages.join("\n\n")
-
-    return prompt
-
-}
-
-
-function formatLlama3Message(role: string, content: string) {
-    
-    let message = `<|start_header_id|>${role}<|end_header_id|>\n`
-    message += content        
-    message += `\n<|eot_id|>`
-
-    return message
-
-}
-
-function formatSchemaInstruction(schema: object) {
-
-    return `You must answer using the following JSONSchema.
-    Do not write anything other than a JSON object corresponding to the schema:
-    <schema>
-    ${JSON.stringify(schema, undefined, 2)}
-    </schema>
-    `
-    
-}