mirror of
https://github.com/zadam/trilium.git
synced 2026-07-08 16:03:17 +02:00
refactor(llm): extract helpers out of tools
This commit is contained in:
Elian Doran
@@ -12,7 +12,7 @@ import yaml from "js-yaml";
|
||||
import becca from "../../../becca/becca.js";
|
||||
import { getSkillsSummary } from "../skills/index.js";
|
||||
import { allToolRegistries } from "../tools/index.js";
|
||||
import { getNoteMeta } from "../tools/note_tools.js";
|
||||
import { getNoteMeta } from "../tools/helpers.js";
|
||||
import type { LlmProvider, LlmProviderConfig, ModelInfo, ModelPricing, StreamResult } from "../types.js";
|
||||
|
||||
const DEFAULT_MAX_TOKENS = 8096;
|
||||
|
||||
141
apps/server/src/services/llm/tools/helpers.ts
Normal file
141
apps/server/src/services/llm/tools/helpers.ts
Normal file
@@ -0,0 +1,141 @@
|
||||
/**
|
||||
* Shared helpers for LLM tools — content conversion, metadata building, and previews.
|
||||
*/
|
||||
|
||||
import type BAttachment from "../../../becca/entities/battachment.js";
|
||||
import type BNote from "../../../becca/entities/bnote.js";
|
||||
import becca from "../../../becca/becca.js";
|
||||
import markdownExport from "../../export/markdown.js";
|
||||
import markdownImport from "../../import/markdown.js";
|
||||
|
||||
const CONTENT_PREVIEW_MAX_LENGTH = 500;
|
||||
const ATTACHMENT_PREVIEW_MAX_LENGTH = 200;
|
||||
|
||||
/**
|
||||
* Convert note content to a format suitable for LLM consumption.
|
||||
* Text notes are converted from HTML to Markdown to reduce token usage.
|
||||
*/
|
||||
export function getNoteContentForLlm(note: { type: string; blobId?: string; getContent: () => string | Buffer }) {
|
||||
const content = note.getContent();
|
||||
if (typeof content !== "string") {
|
||||
// For binary content (images, files), use extracted text if available.
|
||||
const blob = note.blobId ? becca.getBlob({ blobId: note.blobId }) : null;
|
||||
if (blob?.textRepresentation) {
|
||||
return `[extracted text from ${note.type}]\n${blob.textRepresentation}`;
|
||||
}
|
||||
return "[binary content]";
|
||||
}
|
||||
if (note.type === "text") {
|
||||
return markdownExport.toMarkdown(content);
|
||||
}
|
||||
return content;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert LLM-provided content to a format suitable for storage.
|
||||
* For text notes, converts Markdown to HTML.
|
||||
*/
|
||||
export function setNoteContentFromLlm(note: { type: string; title: string; setContent: (content: string) => void }, content: string) {
|
||||
if (note.type === "text") {
|
||||
note.setContent(markdownImport.renderToHtml(content, note.title));
|
||||
} else {
|
||||
note.setContent(content);
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Return a short plain-text content preview for a note, truncated to
|
||||
* {@link CONTENT_PREVIEW_MAX_LENGTH} characters. Useful for giving an LLM a
|
||||
* glimpse of the content without sending the full body.
|
||||
*/
|
||||
export function getContentPreview(note: { type: string; blobId?: string; getContent: () => string | Buffer; isContentAvailable: () => boolean }): string | null {
|
||||
if (!note.isContentAvailable()) {
|
||||
return null;
|
||||
}
|
||||
|
||||
const full = getNoteContentForLlm(note);
|
||||
if (!full || full === "[binary content]") {
|
||||
return null;
|
||||
}
|
||||
|
||||
if (full.length <= CONTENT_PREVIEW_MAX_LENGTH) {
|
||||
return full;
|
||||
}
|
||||
|
||||
return `${full.slice(0, CONTENT_PREVIEW_MAX_LENGTH)}…`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Return a short content preview for an attachment, or null if no readable
|
||||
* content is available. For text attachments the raw content is used; for
|
||||
* binary attachments (PDF, images) the OCR/extracted text is used when present.
|
||||
*/
|
||||
export function getAttachmentContentPreview(att: BAttachment): string | null {
|
||||
let text: string | null = null;
|
||||
|
||||
if (att.hasStringContent()) {
|
||||
const content = att.getContent();
|
||||
text = typeof content === "string" ? content : content.toString("utf-8");
|
||||
} else {
|
||||
const blob = att.blobId ? becca.getBlob({ blobId: att.blobId }) : null;
|
||||
text = blob?.textRepresentation ?? null;
|
||||
}
|
||||
|
||||
if (!text) {
|
||||
return null;
|
||||
}
|
||||
|
||||
if (text.length <= ATTACHMENT_PREVIEW_MAX_LENGTH) {
|
||||
return text;
|
||||
}
|
||||
|
||||
return `${text.slice(0, ATTACHMENT_PREVIEW_MAX_LENGTH)}…`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Build the full metadata object for a note. Used by both the `get_note` tool
|
||||
* and the system prompt.
|
||||
*/
|
||||
export function getNoteMeta(note: BNote) {
|
||||
return {
|
||||
noteId: note.noteId,
|
||||
isProtected: note.isProtected,
|
||||
title: note.title,
|
||||
type: note.type,
|
||||
mime: note.mime,
|
||||
blobId: note.blobId,
|
||||
dateCreated: note.dateCreated,
|
||||
dateModified: note.dateModified,
|
||||
utcDateCreated: note.utcDateCreated,
|
||||
utcDateModified: note.utcDateModified,
|
||||
parentNoteIds: note.getParentNotes().map((p) => p.noteId),
|
||||
childNoteIds: note.getChildNotes().map((ch) => ch.noteId),
|
||||
parentBranchIds: note.getParentBranches().map((p) => p.branchId),
|
||||
childBranchIds: note.getChildBranches().map((ch) => ch.branchId),
|
||||
attributes: note.getAttributes().map((attr) => ({
|
||||
attributeId: attr.attributeId,
|
||||
noteId: attr.noteId,
|
||||
type: attr.type,
|
||||
name: attr.name,
|
||||
value: attr.value,
|
||||
position: attr.position,
|
||||
isInheritable: attr.isInheritable,
|
||||
utcDateModified: attr.utcDateModified
|
||||
})),
|
||||
contentPreview: getContentPreview(note),
|
||||
attachments: note.getAttachments().map((att) => ({
|
||||
attachmentId: att.attachmentId,
|
||||
ownerId: att.ownerId,
|
||||
role: att.role,
|
||||
mime: att.mime,
|
||||
title: att.title,
|
||||
position: att.position,
|
||||
blobId: att.blobId,
|
||||
dateModified: att.dateModified,
|
||||
utcDateModified: att.utcDateModified,
|
||||
utcDateScheduledForErasureSince: att.utcDateScheduledForErasureSince,
|
||||
contentLength: att.contentLength,
|
||||
contentPreview: getAttachmentContentPreview(att)
|
||||
}))
|
||||
};
|
||||
}
|
||||
@@ -4,149 +4,14 @@
|
||||
|
||||
import { z } from "zod";
|
||||
|
||||
import type BAttachment from "../../../becca/entities/battachment.js";
|
||||
import type BNote from "../../../becca/entities/bnote.js";
|
||||
import becca from "../../../becca/becca.js";
|
||||
import markdownExport from "../../export/markdown.js";
|
||||
import markdownImport from "../../import/markdown.js";
|
||||
import noteService from "../../notes.js";
|
||||
import SearchContext from "../../search/search_context.js";
|
||||
import searchService from "../../search/services/search.js";
|
||||
import { getNoteContentForLlm, getNoteMeta, setNoteContentFromLlm } from "./helpers.js";
|
||||
import { defineTools } from "./tool_registry.js";
|
||||
|
||||
const CONTENT_PREVIEW_MAX_LENGTH = 500;
|
||||
|
||||
/**
|
||||
* Return a short plain-text content preview for a note, truncated to
|
||||
* {@link CONTENT_PREVIEW_MAX_LENGTH} characters. Useful for giving an LLM a
|
||||
* glimpse of the content without sending the full body.
|
||||
*/
|
||||
export function getContentPreview(note: { type: string; blobId?: string; getContent: () => string | Buffer; isContentAvailable: () => boolean }): string | null {
|
||||
if (!note.isContentAvailable()) {
|
||||
return null;
|
||||
}
|
||||
|
||||
const full = getNoteContentForLlm(note);
|
||||
if (!full || full === "[binary content]") {
|
||||
return null;
|
||||
}
|
||||
|
||||
if (full.length <= CONTENT_PREVIEW_MAX_LENGTH) {
|
||||
return full;
|
||||
}
|
||||
|
||||
return `${full.slice(0, CONTENT_PREVIEW_MAX_LENGTH)}…`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Build the full metadata object for a note. Used by both the `get_note` tool
|
||||
* and the system prompt.
|
||||
*/
|
||||
export function getNoteMeta(note: BNote) {
|
||||
return {
|
||||
noteId: note.noteId,
|
||||
isProtected: note.isProtected,
|
||||
title: note.title,
|
||||
type: note.type,
|
||||
mime: note.mime,
|
||||
blobId: note.blobId,
|
||||
dateCreated: note.dateCreated,
|
||||
dateModified: note.dateModified,
|
||||
utcDateCreated: note.utcDateCreated,
|
||||
utcDateModified: note.utcDateModified,
|
||||
parentNoteIds: note.getParentNotes().map((p) => p.noteId),
|
||||
childNoteIds: note.getChildNotes().map((ch) => ch.noteId),
|
||||
parentBranchIds: note.getParentBranches().map((p) => p.branchId),
|
||||
childBranchIds: note.getChildBranches().map((ch) => ch.branchId),
|
||||
attributes: note.getAttributes().map((attr) => ({
|
||||
attributeId: attr.attributeId,
|
||||
noteId: attr.noteId,
|
||||
type: attr.type,
|
||||
name: attr.name,
|
||||
value: attr.value,
|
||||
position: attr.position,
|
||||
isInheritable: attr.isInheritable,
|
||||
utcDateModified: attr.utcDateModified
|
||||
})),
|
||||
contentPreview: getContentPreview(note),
|
||||
attachments: note.getAttachments().map((att) => ({
|
||||
attachmentId: att.attachmentId,
|
||||
ownerId: att.ownerId,
|
||||
role: att.role,
|
||||
mime: att.mime,
|
||||
title: att.title,
|
||||
position: att.position,
|
||||
blobId: att.blobId,
|
||||
dateModified: att.dateModified,
|
||||
utcDateModified: att.utcDateModified,
|
||||
utcDateScheduledForErasureSince: att.utcDateScheduledForErasureSince,
|
||||
contentLength: att.contentLength,
|
||||
contentPreview: getAttachmentContentPreview(att)
|
||||
}))
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert note content to a format suitable for LLM consumption.
|
||||
* Text notes are converted from HTML to Markdown to reduce token usage.
|
||||
*/
|
||||
export function getNoteContentForLlm(note: { type: string; blobId?: string; getContent: () => string | Buffer }) {
|
||||
const content = note.getContent();
|
||||
if (typeof content !== "string") {
|
||||
// For binary content (images, files), use extracted text if available.
|
||||
const blob = note.blobId ? becca.getBlob({ blobId: note.blobId }) : null;
|
||||
if (blob?.textRepresentation) {
|
||||
return `[extracted text from ${note.type}]\n${blob.textRepresentation}`;
|
||||
}
|
||||
return "[binary content]";
|
||||
}
|
||||
if (note.type === "text") {
|
||||
return markdownExport.toMarkdown(content);
|
||||
}
|
||||
return content;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert LLM-provided content to a format suitable for storage.
|
||||
* For text notes, converts Markdown to HTML.
|
||||
*/
|
||||
function setNoteContentFromLlm(note: { type: string; title: string; setContent: (content: string) => void }, content: string) {
|
||||
if (note.type === "text") {
|
||||
note.setContent(markdownImport.renderToHtml(content, note.title));
|
||||
} else {
|
||||
note.setContent(content);
|
||||
}
|
||||
}
|
||||
|
||||
const ATTACHMENT_PREVIEW_MAX_LENGTH = 200;
|
||||
|
||||
/**
|
||||
* Return a short content preview for an attachment, or null if no readable
|
||||
* content is available. For text attachments the raw content is used; for
|
||||
* binary attachments (PDF, images) the OCR/extracted text is used when present.
|
||||
*/
|
||||
function getAttachmentContentPreview(att: BAttachment): string | null {
|
||||
let text: string | null = null;
|
||||
|
||||
if (att.hasStringContent()) {
|
||||
const content = att.getContent();
|
||||
text = typeof content === "string" ? content : content.toString("utf-8");
|
||||
} else {
|
||||
const blob = att.blobId ? becca.getBlob({ blobId: att.blobId }) : null;
|
||||
text = blob?.textRepresentation ?? null;
|
||||
}
|
||||
|
||||
if (!text) {
|
||||
return null;
|
||||
}
|
||||
|
||||
if (text.length <= ATTACHMENT_PREVIEW_MAX_LENGTH) {
|
||||
return text;
|
||||
}
|
||||
|
||||
return `${text.slice(0, ATTACHMENT_PREVIEW_MAX_LENGTH)}…`;
|
||||
}
|
||||
|
||||
export const noteTools = defineTools({
|
||||
search_notes: {
|
||||
description: [
|
||||
|
||||
Reference in New Issue
Block a user