node-llama-cpp/dist/chatWrappers/Llama2ChatWrapper.js

Run AI models locally on your machine with node.js bindings for llama.cpp. Enforce a JSON schema on the model output on the generation level

import { ChatWrapper } from "../ChatWrapper.js";
import { SpecialToken, LlamaText, SpecialTokensText } from "../utils/LlamaText.js";
// source: https://huggingface.co/blog/llama2#how-to-prompt-llama-2
export class Llama2ChatWrapper extends ChatWrapper {
    wrapperName = "Llama2Chat";
    /** @internal */ _addSpaceBeforeEos;
    constructor({ addSpaceBeforeEos = false } = {}) {
        super();
        this._addSpaceBeforeEos = addSpaceBeforeEos;
    }
    generateContextState({ chatHistory, availableFunctions, documentFunctionParams }) {
        const historyWithFunctions = this.addAvailableFunctionsSystemMessageToHistory(chatHistory, availableFunctions, {
            documentParams: documentFunctionParams
        });
        const resultItems = [];
        let systemTexts = [];
        let userTexts = [];
        let modelTexts = [];
        let currentAggregateFocus = null;
        function flush() {
            if (systemTexts.length > 0 || userTexts.length > 0 || modelTexts.length > 0)
                resultItems.push({
                    system: LlamaText.joinValues("\n\n", systemTexts),
                    user: LlamaText.joinValues("\n\n", userTexts),
                    model: LlamaText.joinValues("\n\n", modelTexts)
                });
            systemTexts = [];
            userTexts = [];
            modelTexts = [];
        }
        for (const item of historyWithFunctions) {
            if (item.type === "system") {
                if (currentAggregateFocus !== "system")
                    flush();
                currentAggregateFocus = "system";
                systemTexts.push(LlamaText.fromJSON(item.text));
            }
            else if (item.type === "user") {
                if (currentAggregateFocus !== "system" && currentAggregateFocus !== "user")
                    flush();
                currentAggregateFocus = "user";
                userTexts.push(LlamaText(item.text));
            }
            else if (item.type === "model") {
                currentAggregateFocus = "model";
                modelTexts.push(this.generateModelResponseText(item.response));
            }
            else
                void item;
        }
        flush();
        const contextText = LlamaText(resultItems.map(({ system, user, model }, index) => {
            const isLastItem = index === resultItems.length - 1;
            return LlamaText([
                new SpecialToken("BOS"),
                (system.values.length === 0 && user.values.length === 0)
                    ? LlamaText([])
                    : LlamaText([
                        new SpecialTokensText("[INST] "),
                        system.values.length === 0
                            ? LlamaText([])
                            : LlamaText([
                                new SpecialTokensText("<<SYS>>\n"),
                                system,
                                new SpecialTokensText("\n<</SYS>>\n\n")
                            ]),
                        user,
                        new SpecialTokensText(" [/INST] ")
                    ]),
                model,
                this._addSpaceBeforeEos
                    ? " "
                    : "",
                isLastItem
                    ? LlamaText([])
                    : new SpecialToken("EOS")
            ]);
        }));
        return {
            contextText,
            stopGenerationTriggers: [
                LlamaText(new SpecialToken("EOS")),
                LlamaText("</s>")
            ]
        };
    }
    /** @internal */
    static _getOptionConfigurationsToTestIfCanSupersedeJinjaTemplate() {
        return [
            { addSpaceBeforeEos: false },
            { addSpaceBeforeEos: true }
        ];
    }
}
//# sourceMappingURL=Llama2ChatWrapper.js.map