node-llama-cpp 2.8.6 → 3.0.0-beta.10

package/dist/evaluator/LlamaChat/LlamaChat.js

@@ -0,0 +1,705 @@
+import { DisposeAggregator, DisposedError, EventRelay } from "lifecycle-utils";
+import { resolveChatWrapper } from "../../utils/resolveChatWrapper.js";
+import { removeNullFields } from "../../utils/removeNullFields.js";
+import { LlamaGrammarEvaluationState } from "../LlamaGrammarEvaluationState.js";
+import { AbortError } from "../../AbortError.js";
+import { StopGenerationDetector } from "../../utils/StopGenerationDetector.js";
+import { TokenStreamRegulator } from "../../utils/TokenStreamRegulator.js";
+import { eraseFirstResponseAndKeepFirstSystemChatContextShiftStrategy } from "./utils/contextShiftStrategies/eraseFirstResponseAndKeepFirstSystemChatContextShiftStrategy.js";
+import { FunctionCallGrammar, LlamaFunctionCallValidationError } from "./utils/FunctionCallGrammar.js";
+const defaultContextShiftOptions = {
+    size: (sequence) => Math.max(1, Math.floor(sequence.context.contextSize / 10)),
+    strategy: "eraseFirstResponseAndKeepFirstSystem",
+    lastEvaluationMetadata: null
+};
+const UNKNOWN_UNICODE_CHAR = "\ufffd";
+export class LlamaChat {
+    /** @internal */ _chatWrapper;
+    /** @internal */ _disposeAggregator = new DisposeAggregator();
+    /** @internal */ _autoDisposeSequence;
+    /** @internal */ _sequence;
+    onDispose = new EventRelay();
+    constructor({ contextSequence, chatWrapper = "auto", autoDisposeSequence = true }) {
+        if (contextSequence == null)
+            throw new Error("contextSequence cannot be null");
+        if (contextSequence.disposed)
+            throw new DisposedError();
+        this._sequence = contextSequence;
+        this._autoDisposeSequence = autoDisposeSequence;
+        this._disposeAggregator.add(this._sequence.onDispose.createListener(() => {
+            this.dispose();
+        }));
+        this._disposeAggregator.add(this.onDispose.dispatchEvent);
+        this._chatWrapper = resolveChatWrapper(chatWrapper, contextSequence.model);
+    }
+    dispose({ disposeSequence = this._autoDisposeSequence } = {}) {
+        if (this._sequence == null)
+            return;
+        if (disposeSequence)
+            this._sequence.dispose();
+        this._sequence = null;
+        this._disposeAggregator.dispose();
+    }
+    /** @hidden */
+    [Symbol.dispose]() {
+        return this.dispose();
+    }
+    get disposed() {
+        return this._sequence == null;
+    }
+    get chatWrapper() {
+        if (this._sequence == null)
+            throw new DisposedError();
+        return this._chatWrapper;
+    }
+    get sequence() {
+        if (this._sequence == null)
+            throw new DisposedError();
+        return this._sequence;
+    }
+    get context() {
+        return this.sequence.context;
+    }
+    get model() {
+        return this.sequence.model;
+    }
+    async generateResponse(history, { onToken, signal, maxTokens, temperature, minP, topK, topP, grammar, trimWhitespaceSuffix = false, repeatPenalty = {}, evaluationPriority = 5, functions, documentFunctionParams, contextShift = defaultContextShiftOptions, lastEvaluationContextWindow: { history: lastEvaluationContextWindowHistory, minimumOverlapPercentageToPreventContextShift = 0.5 } = {} } = {}) {
+        const functionsEnabled = (functions != null && Object.keys(functions).length > 0);
+        if (grammar != null && functionsEnabled)
+            throw new Error("Using both grammar and functions is not supported yet");
+        if (signal?.aborted)
+            throw new AbortError();
+        if (this._sequence == null)
+            throw new DisposedError();
+        let resolvedHistory = this._sequence.isLoadedToMemory
+            ? history.slice()
+            : history.map(removeRawFromHistoryItem);
+        if (resolvedHistory.length === 0 || resolvedHistory[resolvedHistory.length - 1].type !== "model")
+            resolvedHistory.push({
+                type: "model",
+                response: []
+            });
+        const model = this._sequence.model;
+        const context = this._sequence.context;
+        const eosToken = model.tokens.eos;
+        const resolvedContextShift = {
+            ...defaultContextShiftOptions,
+            ...removeNullFields(contextShift)
+        };
+        const { lastTokens: repeatPenaltyLastTokens = 64, punishTokensFilter, penalizeNewLine, penalty, frequencyPenalty, presencePenalty } = repeatPenalty === false
+            ? { lastTokens: 0 }
+            : repeatPenalty;
+        const lastModelResponse = getLastTextModelResponseFromChatHistory(resolvedHistory);
+        const res = [];
+        const pendingTokens = [];
+        let ignoredStartTextTokens = [];
+        const functionCallTokens = [];
+        const repeatPenaltyEnabled = repeatPenaltyLastTokens > 0;
+        const grammarEvaluationState = grammar != null
+            ? new LlamaGrammarEvaluationState({ grammar })
+            : undefined;
+        let functionsGrammar = functionsEnabled
+            ? new FunctionCallGrammar(model._llama, functions, this._chatWrapper, false)
+            : undefined;
+        let functionsEvaluationState = (functionsEnabled && functionsGrammar != null)
+            ? new LlamaGrammarEvaluationState({
+                grammar: functionsGrammar
+            })
+            : undefined;
+        const streamRegulator = new TokenStreamRegulator();
+        const stopGenerationDetector = new StopGenerationDetector();
+        const functionSyntaxStartDetector = new StopGenerationDetector();
+        const functionSyntaxEndDetector = new StopGenerationDetector();
+        const disengageInitiallyEngagedFunctionMode = new StopGenerationDetector();
+        const ignoreStartTextDetector = new StopGenerationDetector();
+        const locksToReleaseOnValidGeneration = [];
+        const functionCallTokenSyntaxLocks = [];
+        let generatedTokens = 0;
+        let isFirstEvaluation = true;
+        let inFunctionEvaluationMode = false;
+        let initiallyEngagedFunctionMode = false;
+        let lastContextWindowHistory = resolvedHistory;
+        let lastHistoryCompressionMetadata = resolvedContextShift.lastEvaluationMetadata;
+        const ensureNotAborted = () => {
+            if (signal?.aborted)
+                throw new AbortError();
+            if (this._sequence == null)
+                throw new DisposedError();
+        };
+        const getPenaltyTokens = () => {
+            if (this._sequence == null)
+                throw new DisposedError();
+            let punishTokens = res.slice(-repeatPenaltyLastTokens);
+            if (punishTokensFilter != null)
+                punishTokens = punishTokensFilter(punishTokens);
+            if (!penalizeNewLine) {
+                const nlToken = model.tokens.nl;
+                if (nlToken != null)
+                    punishTokens = punishTokens.filter(token => token !== nlToken);
+            }
+            return punishTokens;
+        };
+        const getResolvedHistoryWithCurrentModelResponse = () => {
+            if (res.length === 0)
+                return resolvedHistory;
+            let modelResponse = model.detokenize(res);
+            if (grammar?.trimWhitespaceSuffix || trimWhitespaceSuffix)
+                modelResponse = modelResponse.trimEnd();
+            if (modelResponse === "")
+                return resolvedHistory;
+            return setLastModelTextResponseInChatHistory(resolvedHistory, lastModelResponse + modelResponse);
+        };
+        const removeFoundStartIgnoreTextsFromPendingTokens = () => {
+            if (res.length === 0 && pendingTokens.length > 0) {
+                ignoreStartTextDetector.clearInProgressStops();
+                ignoreStartTextDetector.clearTriggeredStops();
+                let mostExhaustiveTriggeredStops = null;
+                for (let i = 0; i < pendingTokens.length; i++) {
+                    ignoreStartTextDetector.recordGeneration({
+                        text: model.detokenize([pendingTokens[i]]),
+                        tokens: [pendingTokens[i]],
+                        startNewChecks: i === 0
+                    });
+                    if (ignoreStartTextDetector.hasTriggeredStops) {
+                        mostExhaustiveTriggeredStops = ignoreStartTextDetector.getTriggeredStops();
+                        ignoreStartTextDetector.clearTriggeredStops();
+                    }
+                    else if (!ignoreStartTextDetector.hasInProgressStops)
+                        break;
+                }
+                if (mostExhaustiveTriggeredStops != null) {
+                    const [mostExhaustiveTriggeredStop] = mostExhaustiveTriggeredStops;
+                    if (mostExhaustiveTriggeredStop != null) {
+                        ignoredStartTextTokens = mostExhaustiveTriggeredStop.stopTrigger
+                            .map((stopTrigger) => {
+                            if (typeof stopTrigger === "string")
+                                return model.tokenize(stopTrigger);
+                            else
+                                return [stopTrigger];
+                        })
+                            .flat(1);
+                        const newPendingTokens = mostExhaustiveTriggeredStop.remainingGenerations
+                            .map((generation) => {
+                            if (typeof generation === "string")
+                                return model.tokenize(generation);
+                            else
+                                return generation;
+                        })
+                            .flat(1);
+                        pendingTokens.length = 0;
+                        pendingTokens.push(...newPendingTokens);
+                    }
+                }
+            }
+        };
+        if (grammar != null)
+            StopGenerationDetector.resolveStopTriggers(grammar.stopGenerationTriggers, model.tokenize)
+                .map((stopTrigger) => stopGenerationDetector.addStopTrigger(stopTrigger));
+        if (functions != null && Object.keys(functions).length > 0)
+            functionSyntaxStartDetector.addStopTrigger([this._chatWrapper.settings.functions.call.prefix]);
+        // eslint-disable-next-line no-constant-condition
+        while (true) {
+            ensureNotAborted();
+            let shouldContextShift = false;
+            const queuedChunkTokens = streamRegulator.getAllQueuedChunkTokens();
+            const { history: contextWindowHistory, stopGenerationTriggers, tokens: contextWindowTokens, newResolvedHistory, newHistoryCompressionMetadata, ignoreStartText, functionCallInitiallyEngaged, disengageInitiallyEngagedFunctionCall } = await getContextWindow({
+                resolvedHistory: getResolvedHistoryWithCurrentModelResponse(),
+                resolvedContextShift,
+                lastHistoryCompressionMetadata,
+                pendingTokensCount: pendingTokens.length + queuedChunkTokens.length,
+                isFirstEvaluation,
+                chatWrapper: this._chatWrapper,
+                lastEvaluationContextWindowHistory,
+                minimumOverlapPercentageToPreventContextShift,
+                sequence: this._sequence,
+                minFreeContextTokens: 1,
+                functions: functionsEnabled ? functions : undefined,
+                documentFunctionParams
+            });
+            ensureNotAborted();
+            if (generatedTokens === 0) {
+                StopGenerationDetector.resolveStopTriggers(ignoreStartText, model.tokenize)
+                    .map((stopTrigger) => ignoreStartTextDetector.addStopTrigger(stopTrigger));
+                if (functionsEnabled) {
+                    initiallyEngagedFunctionMode = functionCallInitiallyEngaged;
+                    StopGenerationDetector.resolveStopTriggers(disengageInitiallyEngagedFunctionCall, model.tokenize)
+                        .map((stopTrigger) => disengageInitiallyEngagedFunctionMode.addStopTrigger(stopTrigger));
+                    if (initiallyEngagedFunctionMode) {
+                        inFunctionEvaluationMode = true;
+                        functionsGrammar = new FunctionCallGrammar(model._llama, functions, this._chatWrapper, true);
+                        functionsEvaluationState = new LlamaGrammarEvaluationState({
+                            grammar: functionsGrammar
+                        });
+                    }
+                }
+            }
+            const tokens = [...contextWindowTokens, ...ignoredStartTextTokens, ...pendingTokens, ...queuedChunkTokens];
+            resolvedHistory = newResolvedHistory;
+            lastHistoryCompressionMetadata = newHistoryCompressionMetadata;
+            lastContextWindowHistory = contextWindowHistory;
+            const contextWindowLastModelResponse = getLastTextModelResponseFromChatHistory(contextWindowHistory);
+            const contextWindowsRes = [];
+            StopGenerationDetector.resolveStopTriggers(stopGenerationTriggers, model.tokenize)
+                .map((stopTrigger) => stopGenerationDetector.addStopTrigger(stopTrigger));
+            if (functionsGrammar != null)
+                StopGenerationDetector.resolveStopTriggers(functionsGrammar.stopGenerationTriggers, model.tokenize)
+                    .map((stopTrigger) => functionSyntaxEndDetector.addStopTrigger(stopTrigger));
+            let { firstDifferentIndex } = this._sequence.compareContextTokens(tokens);
+            // we need to decode at least one token to generate a response
+            if (firstDifferentIndex === tokens.length && firstDifferentIndex > 0)
+                firstDifferentIndex -= 1;
+            tokens.splice(0, firstDifferentIndex);
+            if (firstDifferentIndex < this._sequence.nextTokenIndex) {
+                await this._sequence.eraseContextTokenRanges([{
+                        start: firstDifferentIndex,
+                        end: this._sequence.nextTokenIndex
+                    }]);
+                ensureNotAborted();
+            }
+            const evaluationIterator = this._sequence.evaluate(tokens, removeNullFields({
+                temperature, minP, topK, topP,
+                grammarEvaluationState: () => {
+                    if (inFunctionEvaluationMode)
+                        return functionsEvaluationState;
+                    return grammarEvaluationState;
+                },
+                repeatPenalty: !repeatPenaltyEnabled ? undefined : {
+                    punishTokens: getPenaltyTokens,
+                    penalty,
+                    frequencyPenalty,
+                    presencePenalty
+                },
+                evaluationPriority,
+                yieldEosToken: true
+            }));
+            for await (const token of evaluationIterator) {
+                ensureNotAborted();
+                generatedTokens++;
+                const tokens = [token];
+                const text = model.detokenize([token]);
+                const queuedTokenRelease = streamRegulator.addChunk({ tokens, text });
+                if (initiallyEngagedFunctionMode)
+                    disengageInitiallyEngagedFunctionMode.recordGeneration({ text, tokens, startNewChecks: generatedTokens === 1 });
+                if (text === UNKNOWN_UNICODE_CHAR || ((grammar?.trimWhitespaceSuffix || trimWhitespaceSuffix) && text.trim() === "")) {
+                    locksToReleaseOnValidGeneration.push(queuedTokenRelease.createTextIndexLock(0));
+                }
+                else {
+                    while (locksToReleaseOnValidGeneration.length > 0)
+                        locksToReleaseOnValidGeneration.shift().dispose();
+                }
+                functionSyntaxStartDetector.recordGeneration({ text, tokens, queuedTokenRelease });
+                if (initiallyEngagedFunctionMode && disengageInitiallyEngagedFunctionMode.hasTriggeredStops) {
+                    initiallyEngagedFunctionMode = false;
+                    let shouldStopFunctionEvaluationMode = !functionSyntaxStartDetector.hasTriggeredStops;
+                    if (!shouldStopFunctionEvaluationMode && functionsEnabled && functionsGrammar != null) {
+                        const functionCallText = model.detokenize([...functionCallTokens, ...tokens]);
+                        try {
+                            const functionName = functionsGrammar.parseFunctionNameFromPartialCall(functionCallText, {
+                                enableInternalBuiltinFunctions: true,
+                                initialFunctionCallEngaged: true
+                            });
+                            const internalBuiltinFunctions = this._chatWrapper.getInternalBuiltinFunctions({ initialFunctionCallEngaged: true });
+                            if (internalBuiltinFunctions[functionName] != null) {
+                                shouldStopFunctionEvaluationMode = true;
+                            }
+                        }
+                        catch (err) {
+                            if (!(err instanceof LlamaFunctionCallValidationError))
+                                throw err;
+                        }
+                    }
+                    if (shouldStopFunctionEvaluationMode) {
+                        inFunctionEvaluationMode = false;
+                        functionsGrammar = new FunctionCallGrammar(model._llama, functions, this._chatWrapper, false);
+                        functionsEvaluationState = new LlamaGrammarEvaluationState({
+                            grammar: functionsGrammar
+                        });
+                        functionCallTokens.length = 0;
+                        while (functionCallTokenSyntaxLocks.length > 0)
+                            functionCallTokenSyntaxLocks.shift().dispose();
+                        functionSyntaxStartDetector.clearInProgressStops();
+                        functionSyntaxStartDetector.clearTriggeredStops();
+                        functionSyntaxEndDetector.clearInProgressStops();
+                        functionSyntaxEndDetector.clearTriggeredStops();
+                    }
+                }
+                if (!inFunctionEvaluationMode && functionsEnabled && functionsGrammar != null &&
+                    functionSyntaxStartDetector.hasTriggeredStops && functionsEvaluationState != null) {
+                    inFunctionEvaluationMode = true;
+                    functionCallTokenSyntaxLocks.push(queuedTokenRelease.createTextIndexLock(0));
+                    stopGenerationDetector.clearTriggeredStops();
+                    stopGenerationDetector.clearInProgressStops();
+                    pendingTokens.push(...streamRegulator.popFreeChunkTokens());
+                    const triggeredStops = functionSyntaxStartDetector.getTriggeredStops();
+                    const partiallyFreeTokens = streamRegulator.getPartiallyFreeChunk();
+                    const queuedTokensBeforeStopTrigger = getQueuedTokensBeforeStopTrigger(triggeredStops, partiallyFreeTokens, model.tokenize);
+                    pendingTokens.push(...queuedTokensBeforeStopTrigger);
+                    const [firstRemainingGenerationAfterStop] = triggeredStops
+                        .map((stopTrigger) => stopTrigger.remainingGenerations)
+                        .filter((remainingGenerations) => remainingGenerations.length > 0)
+                        .flat(1);
+                    const remainingTextAfterStop = (firstRemainingGenerationAfterStop == null || firstRemainingGenerationAfterStop.length === 0)
+                        ? ""
+                        : typeof firstRemainingGenerationAfterStop === "string"
+                            ? firstRemainingGenerationAfterStop
+                            : model.detokenize(firstRemainingGenerationAfterStop);
+                    functionCallTokens.push(...model.tokenize(this._chatWrapper.settings.functions.call.prefix + remainingTextAfterStop));
+                    for (const functionCallToken of functionCallTokens)
+                        context._acceptTokenOnGrammarEvaluationState(functionsEvaluationState, functionCallToken);
+                }
+                else if (inFunctionEvaluationMode) {
+                    functionCallTokens.push(...tokens);
+                    functionCallTokenSyntaxLocks.push(queuedTokenRelease.createTextIndexLock(0));
+                    functionSyntaxEndDetector.recordGeneration({ text, tokens, queuedTokenRelease });
+                }
+                if (inFunctionEvaluationMode && functionSyntaxEndDetector.hasTriggeredStops && functionsGrammar != null) {
+                    const functionCallText = model.detokenize(functionCallTokens);
+                    const functionCall = functionsGrammar.parseFunctionCall(functionCallText);
+                    let modelResponse = model.detokenize(res);
+                    let contextWindowModelResponse = model.detokenize(contextWindowsRes);
+                    if (grammar?.trimWhitespaceSuffix || trimWhitespaceSuffix) {
+                        modelResponse = modelResponse.trimEnd();
+                        contextWindowModelResponse = contextWindowModelResponse.trimEnd();
+                    }
+                    return {
+                        response: modelResponse,
+                        lastEvaluation: {
+                            contextWindow: setLastModelTextResponseInChatHistory(lastContextWindowHistory, contextWindowLastModelResponse + contextWindowModelResponse),
+                            cleanHistory: setLastModelTextResponseInChatHistory(resolvedHistory, lastModelResponse + modelResponse),
+                            contextShiftMetadata: lastHistoryCompressionMetadata
+                        },
+                        // prevent infinite TS type instantiation
+                        functionCall: functionCall,
+                        metadata: {
+                            stopReason: "functionCall"
+                        }
+                    };
+                }
+                if (!inFunctionEvaluationMode)
+                    stopGenerationDetector.recordGeneration({ text, tokens, queuedTokenRelease });
+                pendingTokens.push(...streamRegulator.popFreeChunkTokens());
+                removeFoundStartIgnoreTextsFromPendingTokens();
+                if (stopGenerationDetector.hasTriggeredStops || token === eosToken) {
+                    const triggeredStops = stopGenerationDetector.getTriggeredStops();
+                    const partiallyFreeTokens = streamRegulator.getPartiallyFreeChunk();
+                    const queuedTokensBeforeStopTrigger = getQueuedTokensBeforeStopTrigger(triggeredStops, partiallyFreeTokens, model.tokenize);
+                    pendingTokens.push(...queuedTokensBeforeStopTrigger);
+                    const [firstRemainingGenerationAfterStop] = triggeredStops
+                        .map((stopTrigger) => stopTrigger.remainingGenerations)
+                        .filter((remainingGenerations) => remainingGenerations.length > 0)
+                        .flat(1);
+                    removeFoundStartIgnoreTextsFromPendingTokens();
+                    if (pendingTokens.length > 0)
+                        onToken?.(pendingTokens.slice());
+                    res.push(...pendingTokens);
+                    contextWindowsRes.push(...pendingTokens);
+                    pendingTokens.length = 0;
+                    let modelResponse = model.detokenize(res);
+                    let contextWindowModelResponse = model.detokenize(contextWindowsRes);
+                    if (grammar?.trimWhitespaceSuffix || trimWhitespaceSuffix) {
+                        modelResponse = modelResponse.trimEnd();
+                        contextWindowModelResponse = contextWindowModelResponse.trimEnd();
+                    }
+                    return {
+                        response: modelResponse,
+                        lastEvaluation: {
+                            contextWindow: setLastModelTextResponseInChatHistory(lastContextWindowHistory, contextWindowLastModelResponse + contextWindowModelResponse),
+                            cleanHistory: setLastModelTextResponseInChatHistory(resolvedHistory, lastModelResponse + modelResponse),
+                            contextShiftMetadata: lastHistoryCompressionMetadata
+                        },
+                        metadata: {
+                            remainingGenerationAfterStop: firstRemainingGenerationAfterStop,
+                            stopReason: token === eosToken
+                                ? "eosToken"
+                                : "stopGenerationTrigger"
+                        }
+                    };
+                }
+                const maxTokensTriggered = maxTokens != null && maxTokens > 0 && generatedTokens >= maxTokens;
+                if (res.length === 0) {
+                    ignoreStartTextDetector.clearInProgressStops();
+                    ignoreStartTextDetector.clearTriggeredStops();
+                    ignoreStartTextDetector.recordGeneration({
+                        text: model.detokenize(pendingTokens),
+                        tokens: pendingTokens
+                    });
+                }
+                if (pendingTokens.length > 0 && (maxTokensTriggered || !ignoreStartTextDetector.hasInProgressStops)) {
+                    removeFoundStartIgnoreTextsFromPendingTokens();
+                    if (pendingTokens.length > 0) {
+                        onToken?.(pendingTokens.slice());
+                        res.push(...pendingTokens);
+                        contextWindowsRes.push(...pendingTokens);
+                        pendingTokens.length = 0;
+                    }
+                }
+                if (maxTokensTriggered) {
+                    let modelResponse = model.detokenize(res);
+                    let contextWindowModelResponse = model.detokenize(contextWindowsRes);
+                    if (grammar?.trimWhitespaceSuffix || trimWhitespaceSuffix) {
+                        modelResponse = modelResponse.trimEnd();
+                        contextWindowModelResponse = contextWindowModelResponse.trimEnd();
+                    }
+                    return {
+                        response: modelResponse,
+                        lastEvaluation: {
+                            contextWindow: setLastModelTextResponseInChatHistory(lastContextWindowHistory, contextWindowLastModelResponse + contextWindowModelResponse),
+                            cleanHistory: setLastModelTextResponseInChatHistory(resolvedHistory, lastModelResponse + modelResponse),
+                            contextShiftMetadata: lastHistoryCompressionMetadata
+                        },
+                        metadata: {
+                            stopReason: "maxTokens"
+                        }
+                    };
+                }
+                if (this._sequence.nextTokenIndex >= context.contextSize) {
+                    shouldContextShift = true;
+                    break;
+                }
+            }
+            isFirstEvaluation = false;
+            if (shouldContextShift)
+                continue;
+            break;
+        }
+        throw new Error("The context size is too small to generate a response");
+    }
+}
+function removeRawFromHistoryItem(historyItem) {
+    if (historyItem.type === "model") {
+        const newHistoryItem = { ...historyItem };
+        newHistoryItem.response = newHistoryItem.response.map((item) => {
+            if (typeof item === "string")
+                return item;
+            else
+                return {
+                    ...item,
+                    raw: undefined
+                };
+        });
+        return newHistoryItem;
+    }
+    return historyItem;
+}
+async function compressHistoryToFitContextSize({ history, contextShiftSize, contextShiftStrategy, contextShiftLastEvaluationMetadata, contextSize, tokenizer, chatWrapper, functions, documentFunctionParams }) {
+    function checkIfHistoryFitsContext(history) {
+        const { contextText } = chatWrapper.generateContextText(history, {
+            availableFunctions: functions,
+            documentFunctionParams
+        });
+        const tokens = contextText.tokenize(tokenizer);
+        return tokens.length <= contextSize - contextShiftSize;
+    }
+    if (contextSize - contextShiftSize <= 0)
+        throw new Error(`The context size (${contextSize}) is too small to fit the context shift size (${contextShiftSize})`);
+    if (checkIfHistoryFitsContext(history))
+        return {
+            compressedHistory: history,
+            metadata: null
+        };
+    if (contextShiftStrategy instanceof Function) {
+        try {
+            const { chatHistory, metadata } = await contextShiftStrategy({
+                chatHistory: history,
+                maxTokensCount: contextSize - contextShiftSize,
+                tokenizer,
+                chatWrapper,
+                lastShiftMetadata: contextShiftLastEvaluationMetadata
+            });
+            if (checkIfHistoryFitsContext(chatHistory))
+                return {
+                    compressedHistory: chatHistory,
+                    metadata
+                };
+            console.warn("The provided context shift strategy did not return a history that fits the context size. " +
+                "Using the default strategy instead.");
+        }
+        catch (err) {
+            console.error("The provided context shift strategy threw an error. " +
+                "Using the default strategy instead.", err);
+        }
+    }
+    else if (contextShiftStrategy !== "eraseFirstResponseAndKeepFirstSystem")
+        console.warn(`Unknown context shift strategy "${contextShiftStrategy}". ` +
+            "Using the default strategy instead.");
+    const { chatHistory, metadata } = await eraseFirstResponseAndKeepFirstSystemChatContextShiftStrategy({
+        chatHistory: history,
+        maxTokensCount: contextSize - contextShiftSize,
+        tokenizer,
+        chatWrapper,
+        lastShiftMetadata: contextShiftLastEvaluationMetadata
+    });
+    if (!checkIfHistoryFitsContext(chatHistory))
+        throw new Error("The default context shift strategy did not return a history that fits the context size. " +
+            "This may happen due to the system prompt being too long");
+    return {
+        compressedHistory: chatHistory,
+        metadata
+    };
+}
+function getQueuedTokensBeforeStopTrigger(triggeredStops, partiallyFreeTokens, tokenizer) {
+    if (partiallyFreeTokens.tokens.length === 0 && partiallyFreeTokens.text.length === 0)
+        return [];
+    else if (partiallyFreeTokens.tokens.length !== 0 && partiallyFreeTokens.text.length === 0)
+        return partiallyFreeTokens.tokens;
+    else if (partiallyFreeTokens.tokens.length === 0 && partiallyFreeTokens.text.length !== 0)
+        return tokenizer(partiallyFreeTokens.text);
+    const triggerThatStartsWithStringIndex = triggeredStops.findIndex((trigger) => trigger.stopTrigger.length > 0 && typeof trigger.stopTrigger[0] === "string");
+    const triggerThatStartsWithTokenIndex = triggeredStops.findIndex((trigger) => trigger.stopTrigger.length > 0 && typeof trigger.stopTrigger[0] !== "string");
+    if (triggerThatStartsWithTokenIndex > 0 && triggerThatStartsWithStringIndex < 0)
+        return partiallyFreeTokens.tokens;
+    else if (triggerThatStartsWithStringIndex > 0 && triggerThatStartsWithTokenIndex < 0)
+        return tokenizer(partiallyFreeTokens.text);
+    const stringTokens = tokenizer(partiallyFreeTokens.text);
+    if (stringTokens.length === partiallyFreeTokens.tokens.length &&
+        stringTokens.every((value, index) => value === partiallyFreeTokens.tokens[index]))
+        return stringTokens;
+    else if (triggerThatStartsWithStringIndex < triggerThatStartsWithTokenIndex)
+        return stringTokens;
+    return partiallyFreeTokens.tokens;
+}
+function getLastTextModelResponseFromChatHistory(chatHistory) {
+    if (chatHistory.length === 0 || chatHistory[chatHistory.length - 1].type !== "model")
+        return "";
+    const lastModelResponseItem = chatHistory[chatHistory.length - 1];
+    const modelResponse = lastModelResponseItem.response;
+    if (modelResponse.length > 0 && typeof modelResponse[modelResponse.length - 1] === "string")
+        return modelResponse[modelResponse.length - 1];
+    return "";
+}
+function setLastModelTextResponseInChatHistory(chatHistory, textResponse) {
+    const newChatHistory = chatHistory.slice();
+    if (newChatHistory.length === 0 || newChatHistory[newChatHistory.length - 1].type !== "model")
+        newChatHistory.push({
+            type: "model",
+            response: []
+        });
+    const lastModelResponseItem = newChatHistory[newChatHistory.length - 1];
+    const newLastModelResponseItem = { ...lastModelResponseItem };
+    newChatHistory[newChatHistory.length - 1] = newLastModelResponseItem;
+    const modelResponse = newLastModelResponseItem.response.slice();
+    newLastModelResponseItem.response = modelResponse;
+    if (modelResponse.length > 0 && typeof modelResponse[modelResponse.length - 1] === "string") {
+        if (textResponse === "")
+            modelResponse.pop();
+        else
+            modelResponse[modelResponse.length - 1] = textResponse;
+    }
+    else if (textResponse !== "")
+        modelResponse.push(textResponse);
+    return newChatHistory;
+}
+async function getContextWindow({ resolvedHistory, resolvedContextShift, lastHistoryCompressionMetadata, pendingTokensCount = 0, isFirstEvaluation, chatWrapper, lastEvaluationContextWindowHistory, minimumOverlapPercentageToPreventContextShift, sequence, minFreeContextTokens = 1, functions, documentFunctionParams }) {
+    if (sequence == null)
+        throw new DisposedError();
+    const model = sequence.model;
+    const context = sequence.context;
+    if (isFirstEvaluation && lastEvaluationContextWindowHistory != null && sequence.isLoadedToMemory) {
+        const newContextWindow = lastEvaluationContextWindowHistory.slice();
+        if (newContextWindow.length === 0 || newContextWindow[newContextWindow.length - 1].type !== "model")
+            newContextWindow.push({
+                type: "model",
+                response: []
+            });
+        const { contextText, stopGenerationTriggers, ignoreStartText, functionCall } = chatWrapper.generateContextText(newContextWindow, {
+            availableFunctions: functions,
+            documentFunctionParams
+        });
+        const tokens = contextText.tokenize(model.tokenize);
+        if (tokens.length + pendingTokensCount + minFreeContextTokens < context.contextSize) {
+            const { firstDifferentIndex } = sequence.compareContextTokens(tokens);
+            const existingEvaluationPercentage = firstDifferentIndex / tokens.length;
+            if (existingEvaluationPercentage >= minimumOverlapPercentageToPreventContextShift)
+                return {
+                    history: newContextWindow,
+                    stopGenerationTriggers,
+                    tokens,
+                    newResolvedHistory: resolvedHistory,
+                    newHistoryCompressionMetadata: lastHistoryCompressionMetadata,
+                    ignoreStartText: ignoreStartText ?? [],
+                    functionCallInitiallyEngaged: functionCall?.initiallyEngaged ?? false,
+                    disengageInitiallyEngagedFunctionCall: functionCall?.disengageInitiallyEngaged ?? []
+                };
+        }
+    }
+    resolvedHistory = sequence.isLoadedToMemory
+        ? resolvedHistory.slice()
+        : resolvedHistory.map(removeRawFromHistoryItem);
+    if (resolvedContextShift.lastEvaluationMetadata != null) {
+        const contextShiftSize = resolvedContextShift.size instanceof Function
+            ? await resolvedContextShift.size(sequence)
+            : resolvedContextShift.size;
+        const { compressedHistory, metadata } = await compressHistoryToFitContextSize({
+            history: resolvedHistory,
+            contextShiftSize: Math.max(contextShiftSize, minFreeContextTokens) + pendingTokensCount,
+            contextShiftStrategy: resolvedContextShift.strategy,
+            contextShiftLastEvaluationMetadata: resolvedContextShift.lastEvaluationMetadata,
+            contextSize: context.contextSize,
+            tokenizer: model.tokenize,
+            chatWrapper: chatWrapper,
+            functions,
+            documentFunctionParams
+        });
+        const { contextText, stopGenerationTriggers, ignoreStartText, functionCall } = chatWrapper.generateContextText(compressedHistory, {
+            availableFunctions: functions,
+            documentFunctionParams
+        });
+        return {
+            history: compressedHistory,
+            stopGenerationTriggers,
+            tokens: contextText.tokenize(model.tokenize),
+            newResolvedHistory: resolvedHistory,
+            newHistoryCompressionMetadata: metadata,
+            ignoreStartText: ignoreStartText ?? [],
+            functionCallInitiallyEngaged: functionCall?.initiallyEngaged ?? false,
+            disengageInitiallyEngagedFunctionCall: functionCall?.disengageInitiallyEngaged ?? []
+        };
+    }
+    {
+        const { contextText, stopGenerationTriggers, ignoreStartText, functionCall } = chatWrapper.generateContextText(resolvedHistory, {
+            availableFunctions: functions,
+            documentFunctionParams
+        });
+        const tokens = contextText.tokenize(model.tokenize);
+        if (tokens.length + pendingTokensCount + minFreeContextTokens < context.contextSize)
+            return {
+                history: resolvedHistory,
+                stopGenerationTriggers,
+                tokens,
+                newResolvedHistory: resolvedHistory,
+                newHistoryCompressionMetadata: lastHistoryCompressionMetadata,
+                ignoreStartText: ignoreStartText ?? [],
+                functionCallInitiallyEngaged: functionCall?.initiallyEngaged ?? false,
+                disengageInitiallyEngagedFunctionCall: functionCall?.disengageInitiallyEngaged ?? []
+            };
+    }
+    const contextShiftSize = resolvedContextShift.size instanceof Function
+        ? await resolvedContextShift.size(sequence)
+        : resolvedContextShift.size;
+    const { compressedHistory, metadata } = await compressHistoryToFitContextSize({
+        history: resolvedHistory,
+        contextShiftSize: Math.max(contextShiftSize, minFreeContextTokens) + pendingTokensCount,
+        contextShiftStrategy: resolvedContextShift.strategy,
+        contextShiftLastEvaluationMetadata: resolvedContextShift.lastEvaluationMetadata,
+        contextSize: context.contextSize,
+        tokenizer: model.tokenize,
+        chatWrapper: chatWrapper,
+        functions,
+        documentFunctionParams
+    });
+    const { contextText, stopGenerationTriggers, ignoreStartText, functionCall } = chatWrapper.generateContextText(compressedHistory, {
+        availableFunctions: functions,
+        documentFunctionParams
+    });
+    return {
+        history: compressedHistory,
+        stopGenerationTriggers,
+        tokens: contextText.tokenize(model.tokenize),
+        newResolvedHistory: resolvedHistory,
+        newHistoryCompressionMetadata: metadata,
+        ignoreStartText: ignoreStartText ?? [],
+        functionCallInitiallyEngaged: functionCall?.initiallyEngaged ?? false,
+        disengageInitiallyEngagedFunctionCall: functionCall?.disengageInitiallyEngaged ?? []
+    };
+}
+//# sourceMappingURL=LlamaChat.js.map