anything-llm/server/utils/AiProviders/ollama/index.js

210 lines
6.2 KiB
JavaScript
Raw Normal View History

const { StringOutputParser } = require("@langchain/core/output_parsers");
const {
writeResponseChunk,
clientAbortedHandler,
} = require("../../helpers/chat/responses");
const { NativeEmbedder } = require("../../EmbeddingEngines/native");
// Docs: https://github.com/jmorganca/ollama/blob/main/docs/api.md
class OllamaAILLM {
constructor(embedder = null, modelPreference = null) {
if (!process.env.OLLAMA_BASE_PATH)
throw new Error("No Ollama Base Path was set.");
this.basePath = process.env.OLLAMA_BASE_PATH;
this.model = modelPreference || process.env.OLLAMA_MODEL_PREF;
this.limits = {
history: this.promptWindowLimit() * 0.15,
system: this.promptWindowLimit() * 0.15,
user: this.promptWindowLimit() * 0.7,
};
this.embedder = embedder ?? new NativeEmbedder();
this.defaultTemp = 0.7;
}
#ollamaClient({ temperature = 0.07 }) {
const { ChatOllama } = require("@langchain/community/chat_models/ollama");
return new ChatOllama({
baseUrl: this.basePath,
model: this.model,
2024-04-02 19:43:04 +02:00
useMLock: true,
temperature,
});
}
// For streaming we use Langchain's wrapper to handle weird chunks
// or otherwise absorb headaches that can arise from Ollama models
#convertToLangchainPrototypes(chats = []) {
const {
HumanMessage,
SystemMessage,
AIMessage,
} = require("@langchain/core/messages");
const langchainChats = [];
const roleToMessageMap = {
system: SystemMessage,
user: HumanMessage,
assistant: AIMessage,
};
for (const chat of chats) {
if (!roleToMessageMap.hasOwnProperty(chat.role)) continue;
const MessageClass = roleToMessageMap[chat.role];
langchainChats.push(new MessageClass({ content: chat.content }));
}
return langchainChats;
}
#appendContext(contextTexts = []) {
if (!contextTexts || !contextTexts.length) return "";
return (
"\nContext:\n" +
contextTexts
.map((text, i) => {
return `[CONTEXT ${i}]:\n${text}\n[END CONTEXT ${i}]\n\n`;
})
.join("")
);
}
streamingEnabled() {
return "streamGetChatCompletion" in this;
}
// Ensure the user set a value for the token limit
// and if undefined - assume 4096 window.
promptWindowLimit() {
const limit = process.env.OLLAMA_MODEL_TOKEN_LIMIT || 4096;
if (!limit || isNaN(Number(limit)))
throw new Error("No Ollama token context limit was set.");
return Number(limit);
}
async isValidChatCompletionModel(_ = "") {
return true;
}
constructPrompt({
systemPrompt = "",
contextTexts = [],
chatHistory = [],
userPrompt = "",
}) {
const prompt = {
role: "system",
content: `${systemPrompt}${this.#appendContext(contextTexts)}`,
};
return [prompt, ...chatHistory, { role: "user", content: userPrompt }];
}
async isSafe(_input = "") {
// Not implemented so must be stubbed
return { safe: true, reasons: [] };
}
async getChatCompletion(messages = null, { temperature = 0.7 }) {
const model = this.#ollamaClient({ temperature });
const textResponse = await model
.pipe(new StringOutputParser())
.invoke(this.#convertToLangchainPrototypes(messages))
.catch((e) => {
throw new Error(
`Ollama::getChatCompletion failed to communicate with Ollama. ${e.message}`
);
});
if (!textResponse || !textResponse.length)
throw new Error(`Ollama::getChatCompletion text response was empty.`);
return textResponse;
}
async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
const model = this.#ollamaClient({ temperature });
const stream = await model
.pipe(new StringOutputParser())
.stream(this.#convertToLangchainPrototypes(messages));
return stream;
}
handleStream(response, stream, responseProps) {
const { uuid = uuidv4(), sources = [] } = responseProps;
return new Promise(async (resolve) => {
let fullText = "";
// Establish listener to early-abort a streaming response
// in case things go sideways or the user does not like the response.
// We preserve the generated text but continue as if chat was completed
// to preserve previously generated content.
const handleAbort = () => clientAbortedHandler(resolve, fullText);
response.on("close", handleAbort);
try {
for await (const chunk of stream) {
if (chunk === undefined)
throw new Error(
"Stream returned undefined chunk. Aborting reply - check model provider logs."
);
const content = chunk.hasOwnProperty("content")
? chunk.content
: chunk;
fullText += content;
writeResponseChunk(response, {
uuid,
sources: [],
type: "textResponseChunk",
textResponse: content,
close: false,
error: false,
});
}
writeResponseChunk(response, {
uuid,
sources,
type: "textResponseChunk",
textResponse: "",
close: true,
error: false,
});
response.removeListener("close", handleAbort);
resolve(fullText);
} catch (error) {
writeResponseChunk(response, {
uuid,
sources: [],
type: "textResponseChunk",
textResponse: "",
close: true,
error: `Ollama:streaming - could not stream chat. ${
error?.cause ?? error.message
}`,
});
response.removeListener("close", handleAbort);
}
});
}
// Simple wrapper for dynamic embedder & normalize interface for all LLM implementations
async embedTextInput(textInput) {
return await this.embedder.embedTextInput(textInput);
}
async embedChunks(textChunks = []) {
return await this.embedder.embedChunks(textChunks);
}
async compressMessages(promptArgs = {}, rawHistory = []) {
const { messageArrayCompressor } = require("../../helpers/chat");
const messageArray = this.constructPrompt(promptArgs);
return await messageArrayCompressor(this, messageArray, rawHistory);
}
}
module.exports = {
OllamaAILLM,
};