@n8n/n8n-nodes-langchain
Version:
119 lines • 4.38 kB
JavaScript
;
Object.defineProperty(exports, "__esModule", { value: true });
exports.NaiveJsonOutputParser = void 0;
exports.isModelWithResponseFormat = isModelWithResponseFormat;
exports.isModelInThinkingMode = isModelInThinkingMode;
exports.isModelWithFormat = isModelWithFormat;
exports.getOutputParserForLLM = getOutputParserForLLM;
exports.executeChain = executeChain;
const output_parsers_1 = require("@langchain/core/output_parsers");
const ai_utilities_1 = require("@n8n/ai-utilities");
const tracing_1 = require("../../../../utils/tracing");
const promptUtils_1 = require("./promptUtils");
class NaiveJsonOutputParser extends output_parsers_1.JsonOutputParser {
async parse(text) {
try {
const directParsed = JSON.parse(text);
return directParsed;
}
catch (e) {
return await super.parse(text);
}
}
}
exports.NaiveJsonOutputParser = NaiveJsonOutputParser;
function isModelWithResponseFormat(llm) {
return ('modelKwargs' in llm &&
!!llm.modelKwargs &&
typeof llm.modelKwargs === 'object' &&
'response_format' in llm.modelKwargs);
}
function isModelInThinkingMode(llm) {
return ('lc_kwargs' in llm &&
'invocationKwargs' in llm.lc_kwargs &&
typeof llm.lc_kwargs.invocationKwargs === 'object' &&
'thinking' in llm.lc_kwargs.invocationKwargs &&
llm.lc_kwargs.invocationKwargs.thinking.type === 'enabled');
}
function isModelWithFormat(llm) {
return 'format' in llm && typeof llm.format !== 'undefined';
}
function getOutputParserForLLM(llm) {
if (isModelWithResponseFormat(llm) && llm.modelKwargs?.response_format?.type === 'json_object') {
return new NaiveJsonOutputParser();
}
if (isModelWithFormat(llm) && llm.format === 'json') {
return new NaiveJsonOutputParser();
}
if (isModelInThinkingMode(llm)) {
return new NaiveJsonOutputParser();
}
if (llm.metadata?.output_format === 'json') {
return new NaiveJsonOutputParser();
}
return new output_parsers_1.StringOutputParser();
}
async function executeSimpleChain({ context, llm, query, prompt, }) {
const outputParser = getOutputParserForLLM(llm);
const chain = prompt.pipe(llm).pipe(outputParser).withConfig((0, tracing_1.getTracingConfig)(context));
const response = await chain.invoke({ query }, { signal: context.getExecutionCancelSignal() });
return [response];
}
function withBuiltInTools(llm) {
const modelTools = llm.metadata?.tools ?? [];
if (modelTools.length && (0, ai_utilities_1.isChatInstance)(llm) && llm.bindTools) {
return llm.bindTools(modelTools);
}
return llm;
}
function prepareLlm(llm, fallbackLlm) {
const mainLlm = withBuiltInTools(llm);
if (fallbackLlm) {
return mainLlm.withFallbacks([withBuiltInTools(fallbackLlm)]);
}
return mainLlm;
}
async function executeChain({ context, itemIndex, query, llm, outputParser, messages, fallbackLlm, }) {
const version = context.getNode().typeVersion;
const model = prepareLlm(llm, fallbackLlm);
if (!outputParser) {
const promptTemplate = await (0, promptUtils_1.createPromptTemplate)({
context,
itemIndex,
llm,
messages,
query,
});
return await executeSimpleChain({
context,
llm: model,
query,
prompt: promptTemplate,
});
}
const formatInstructions = outputParser.getFormatInstructions();
const promptWithInstructions = await (0, promptUtils_1.createPromptTemplate)({
context,
itemIndex,
llm,
messages,
formatInstructions,
query,
});
let chain;
if (version >= 1.9) {
chain = promptWithInstructions
.pipe(model)
.pipe((0, promptUtils_1.getAgentStepsParser)(outputParser))
.withConfig((0, tracing_1.getTracingConfig)(context));
}
else {
chain = promptWithInstructions
.pipe(model)
.pipe(outputParser)
.withConfig((0, tracing_1.getTracingConfig)(context));
}
const response = await chain.invoke({ query }, { signal: context.getExecutionCancelSignal() });
return Array.isArray(response) ? response : [response];
}
//# sourceMappingURL=chainExecutor.js.map