@khoinguyen2002/doc-mcp 1.0.4 → 1.0.5

package/src/db/syncState.ts

@@ -0,0 +1,87 @@
+import { Redis } from "@upstash/redis";
+import { config } from "../config.js";
+
+const HASH_KEY = "doc_sync_state";
+
+let _redis: Redis | null = null;
+
+function getRedis(): Redis {
+  if (!_redis) {
+    _redis = new Redis({
+      url: config.UPSTASH_REDIS_REST_URL,
+      token: config.UPSTASH_REDIS_REST_TOKEN,
+    });
+  }
+  return _redis;
+}
+
+export interface SyncEntry {
+  modifiedTime: string;
+  blockCount: number;
+  title: string;
+}
+
+export async function getAllSyncEntries(): Promise<Record<string, SyncEntry>> {
+  const redis = getRedis();
+  const raw = await redis.hgetall(HASH_KEY);
+  if (!raw) return {};
+
+  const result: Record<string, SyncEntry> = {};
+  for (const [fileId, value] of Object.entries(raw)) {
+    if (!value) continue;
+    try {
+      result[fileId] =
+        typeof value === "string"
+          ? (JSON.parse(value) as SyncEntry)
+          : (value as unknown as SyncEntry);
+    } catch {
+      // skip malformed entries
+    }
+  }
+  return result;
+}
+
+export async function getSyncEntry(fileId: string): Promise<SyncEntry | null> {
+  const redis = getRedis();
+  const raw = await redis.hget(HASH_KEY, fileId);
+  if (!raw) return null;
+  try {
+    return typeof raw === "string"
+      ? (JSON.parse(raw) as SyncEntry)
+      : (raw as unknown as SyncEntry);
+  } catch {
+    return null;
+  }
+}
+
+export async function setSyncEntry(
+  fileId: string,
+  entry: SyncEntry
+): Promise<void> {
+  const redis = getRedis();
+  await redis.hset(HASH_KEY, { [fileId]: JSON.stringify(entry) });
+}
+
+export async function deleteSyncEntry(fileId: string): Promise<void> {
+  const redis = getRedis();
+  await redis.hdel(HASH_KEY, fileId);
+}
+
+// ─── Image Description Cache ──────────────────────────────────────────────────
+// Global hash: md5(imageBinary) → description text
+// Deduplicates across docs (same image used in multiple files reuses description)
+const IMG_DESC_KEY = "img_desc";
+
+export async function getImageDesc(imageHash: string): Promise<string | null> {
+  const redis = getRedis();
+  const raw = await redis.hget(IMG_DESC_KEY, imageHash);
+  return raw ? String(raw) : null;
+}
+
+export async function setImageDesc(
+  imageHash: string,
+  description: string
+): Promise<void> {
+  const redis = getRedis();
+  await redis.hset(IMG_DESC_KEY, { [imageHash]: description });
+}

package/src/db/vector.ts

@@ -1,9 +1,20 @@
-import { QdrantClient } from '@qdrant/js-client-rest';
-import { v4 as uuidv4 } from 'uuid';
-import { config } from '../config.js';
+import { QdrantClient } from "@qdrant/js-client-rest";
+import { v4 as uuidv4, v5 as uuidv5 } from "uuid";
+import { config } from "../config.js";
 
 let client: QdrantClient | null = null;
-const COLLECTION_NAME = 'project_memory';
+const COLLECTION_NAME = "project_memory";
+
+// Fixed namespace for deterministic point IDs (uuid v5)
+const POINT_NAMESPACE = "1b671a64-40d5-491e-99b0-da01ff1f3341";
+
+/**
+ * Deterministic Qdrant point ID: uuidv5(fileId:blockIndex, NS)
+ * Same input → same ID → upsert overwrites correctly.
+ */
+export function getBlockPointId(fileId: string, blockIndex: number): string {
+  return uuidv5(`${fileId}:${blockIndex}`, POINT_NAMESPACE);
+}
 
 export async function initVectorDB() {
   if (!client) {
@@ -13,119 +24,282 @@ export async function initVectorDB() {
     });
     console.error(`Connected to Qdrant at ${config.QDRANT_URL}`);
 
-    // Check if collection exists
     const res = await client.getCollections();
-    const exists = res.collections.some(c => c.name === COLLECTION_NAME);
+    const exists = res.collections.some((c) => c.name === COLLECTION_NAME);
     if (!exists) {
       console.error(`Creating Qdrant collection: ${COLLECTION_NAME}`);
       const dummyVector = await embedText("test");
       const dimension = dummyVector.length;
 
       await client.createCollection(COLLECTION_NAME, {
-        vectors: {
-          size: dimension,
-          distance: "Cosine",
-        },
+        vectors: { size: dimension, distance: "Cosine" },
       });
       await client.createPayloadIndex(COLLECTION_NAME, {
-        field_name: "folderId",
+        field_name: "source",
         field_schema: "keyword",
       });
       await client.createPayloadIndex(COLLECTION_NAME, {
-        field_name: "file_id",
-        field_schema: "keyword",
+        field_name: "block_index",
+        field_schema: "integer",
       });
       await client.createPayloadIndex(COLLECTION_NAME, {
-        field_name: "source",
+        field_name: "block_hash",
         field_schema: "keyword",
       });
-      console.error(`Collection ${COLLECTION_NAME} created with dimension ${dimension}.`);
+      // Full-text index on `text` payload for exact/keyword search.
+      // whitespace tokenizer keeps API paths (e.g. /v1/foo/bar) as single tokens.
+      // lowercase=true makes searches case-insensitive.
+      await client.createPayloadIndex(COLLECTION_NAME, {
+        field_name: "text",
+        field_schema: {
+          type: "text",
+          tokenizer: "whitespace",
+          min_token_len: 2,
+          max_token_len: 200,
+          lowercase: true,
+        } as any,
+      });
+      console.error(
+        `Collection ${COLLECTION_NAME} created with dimension ${dimension}.`
+      );
     }
   }
 }
 
-export async function embedText(text: string): Promise<number[]> {
-  const response = await fetch("https://openrouter.ai/api/v1/embeddings", {
-    method: "POST",
-    headers: {
-      "Authorization": `Bearer ${config.OPENROUTER_API_KEY}`,
-      "Content-Type": "application/json"
-    },
-    body: JSON.stringify({
-      model: config.EMBEDDING_MODEL_ID,
-      input: text
-    })
-  });
+export async function embedText(
+  text: string,
+  maxRetries = 5
+): Promise<number[]> {
+  for (let attempt = 0; attempt < maxRetries; attempt++) {
+    try {
+      const response = await fetch("https://openrouter.ai/api/v1/embeddings", {
+        method: "POST",
+        headers: {
+          Authorization: `Bearer ${config.OPENROUTER_API_KEY}`,
+          "Content-Type": "application/json",
+        },
+        body: JSON.stringify({
+          model: config.EMBEDDING_MODEL_ID,
+          input: text,
+        }),
+      });
+
+      if (!response.ok) {
+        if (response.status === 429 && attempt < maxRetries - 1) {
+          const delay = Math.pow(2, attempt) * 1000 + Math.random() * 1000;
+          console.error(
+            `[Rate Limit] OpenRouter 429. Retrying in ${Math.round(delay)}ms... (Attempt ${attempt + 1}/${maxRetries})`
+          );
+          await new Promise((res) => setTimeout(res, delay));
+          continue;
+        }
+        const errText = await response.text();
+        throw new Error(
+          `OpenRouter Embedding API failed: ${response.status} ${errText}`
+        );
+      }
 
-  if (!response.ok) {
-    const errText = await response.text();
-    throw new Error(`OpenRouter Embedding API failed: ${response.status} ${errText}`);
+      const json: any = await response.json();
+      if (!json.data || !json.data[0] || !json.data[0].embedding) {
+        throw new Error(
+          `Invalid response from OpenRouter: ${JSON.stringify(json)}`
+        );
+      }
+      return json.data[0].embedding;
+    } catch (err: any) {
+      if (attempt >= maxRetries - 1) throw err;
+      const delay = Math.pow(2, attempt) * 1000 + Math.random() * 1000;
+      console.error(
+        `[Error] ${err.message}. Retrying in ${Math.round(delay)}ms... (Attempt ${attempt + 1}/${maxRetries})`
+      );
+      await new Promise((res) => setTimeout(res, delay));
+    }
   }
+  throw new Error("Max retries reached for embedding");
+}
+
+/**
+ * Embed nhiều texts trong 1 API call (batch).
+ * OpenRouter hỗ trợ input: string[] → trả data[i].embedding.
+ */
+export async function embedBatch(
+  texts: string[],
+  maxRetries = 5
+): Promise<number[][]> {
+  if (texts.length === 0) return [];
+
+  for (let attempt = 0; attempt < maxRetries; attempt++) {
+    try {
+      const response = await fetch("https://openrouter.ai/api/v1/embeddings", {
+        method: "POST",
+        headers: {
+          Authorization: `Bearer ${config.OPENROUTER_API_KEY}`,
+          "Content-Type": "application/json",
+        },
+        body: JSON.stringify({
+          model: config.EMBEDDING_MODEL_ID,
+          input: texts,
+        }),
+      });
 
-  const json: any = await response.json();
-  if (!json.data || !json.data[0] || !json.data[0].embedding) {
-    throw new Error("Invalid response from OpenRouter Embedding API");
+      if (!response.ok) {
+        if (response.status === 429 && attempt < maxRetries - 1) {
+          const delay = Math.pow(2, attempt) * 1000 + Math.random() * 1000;
+          console.error(
+            `[Rate Limit] OpenRouter 429 (batch). Retrying in ${Math.round(delay)}ms... (Attempt ${attempt + 1}/${maxRetries})`
+          );
+          await new Promise((res) => setTimeout(res, delay));
+          continue;
+        }
+        const errText = await response.text();
+        throw new Error(
+          `OpenRouter Batch Embedding API failed: ${response.status} ${errText}`
+        );
+      }
+
+      const json: any = await response.json();
+      if (!json.data || !Array.isArray(json.data)) {
+        throw new Error(
+          `Invalid batch response from OpenRouter: ${JSON.stringify(json)}`
+        );
+      }
+      return json.data.map((item: any) => item.embedding);
+    } catch (err: any) {
+      if (attempt >= maxRetries - 1) throw err;
+      const delay = Math.pow(2, attempt) * 1000 + Math.random() * 1000;
+      console.error(
+        `[Error] ${err.message}. Retrying in ${Math.round(delay)}ms... (Attempt ${attempt + 1}/${maxRetries})`
+      );
+      await new Promise((res) => setTimeout(res, delay));
+    }
   }
+  throw new Error("Max retries reached for batch embedding");
+}
 
-  return json.data[0].embedding;
+export interface ChunkUpsert {
+  pointId: string;
+  vector: number[];
+  text: string;
+  title: string;
+  blockIndex: number;
+  blockHash: string;
+  source: string;
+  offset: number; // character offset in the Markdown string
 }
 
-export async function upsertProjectDocument(folderId: string, text: string, metadata: Record<string, any> = {}): Promise<void> {
+/**
+ * Bulk upsert nhiều chunks vào Qdrant trong 1 HTTP call.
+ */
+export async function upsertChunkBatch(chunks: ChunkUpsert[]): Promise<void> {
   await initVectorDB();
   if (!client) throw new Error("Qdrant not initialized");
-
-  const vector = await embedText(text);
+  if (chunks.length === 0) return;
 
   await client.upsert(COLLECTION_NAME, {
     wait: true,
-    points: [
-      {
-        id: uuidv4(),
-        vector: vector,
-        payload: {
-          folderId,
-          text,
-          source: metadata.source || "user",
-          file_id: metadata.file_id || null,
-          modified_time: metadata.modified_time || null,
-          metadata: JSON.stringify(metadata),
-          createdAt: new Date().toISOString()
-        }
-      }
-    ]
+    points: chunks.map((c) => ({
+      id: c.pointId,
+      vector: c.vector,
+      payload: {
+        text: c.text,
+        title: c.title,
+        block_index: c.blockIndex,
+        block_hash: c.blockHash,
+        source: c.source,
+        offset: c.offset,
+      },
+    })),
+  });
+  console.error(`Upserted ${chunks.length} chunk(s) to Qdrant.`);
+}
+
+/**
+ * Fetch block_hash AND offset for a list of point IDs.
+ * Used to diff block-level changes during re-sync (hash) and
+ * detect stale offsets in unchanged blocks (offset).
+ */
+export async function getBlockMetaByIds(
+  pointIds: string[]
+): Promise<Record<string, { hash: string; offset: number }>> {
+  await initVectorDB();
+  if (!client || pointIds.length === 0) return {};
+
+  const results = await client.retrieve(COLLECTION_NAME, {
+    ids: pointIds,
+    with_payload: ["block_hash", "offset"],
+    with_vector: false,
   });
 
-  console.error(`Upserted document chunk for folder ${folderId}`);
+  const metaMap: Record<string, { hash: string; offset: number }> = {};
+  for (const point of results) {
+    const hash = point.payload?.block_hash as string | undefined;
+    const offset = point.payload?.offset as number | undefined;
+    if (hash !== undefined) {
+      metaMap[point.id as string] = { hash, offset: offset ?? 0 };
+    }
+  }
+  return metaMap;
+}
+
+/**
+ * Update only the `offset` payload field for a set of points (no re-embedding).
+ * Called for unchanged blocks whose character position shifted due to earlier edits.
+ * Uses parallel setPayload calls (lightweight metadata-only updates).
+ */
+export async function updateBlockOffsets(
+  updates: { pointId: string; offset: number }[]
+): Promise<void> {
+  if (updates.length === 0) return;
+  await initVectorDB();
+  if (!client) throw new Error("Qdrant not initialized");
+
+  await Promise.all(
+    updates.map(({ pointId, offset }) =>
+      client!.setPayload(COLLECTION_NAME, {
+        payload: { offset },
+        points: [pointId],
+        wait: false, // fire-and-forget per point; all resolve before function returns
+      })
+    )
+  );
+  console.error(`[Sync] Updated offset for ${updates.length} unchanged block(s).`);
+}
+
+/**
+ * Xóa Qdrant points theo danh sách IDs.
+ */
+export async function deletePointsByIds(pointIds: string[]): Promise<void> {
+  await initVectorDB();
+  if (!client || pointIds.length === 0) return;
+
+  await client.delete(COLLECTION_NAME, {
+    wait: true,
+    points: pointIds,
+  });
+  console.error(`Deleted ${pointIds.length} obsolete block(s) from Qdrant.`);
 }
 
-export async function searchProjectMemory(folderId: string, query: string, topK: number = 3): Promise<any[]> {
+/**
+ * Global semantic search — không filter theo folder hay file.
+ */
+export async function searchProjectMemory(
+  query: string,
+  topK: number = 3
+): Promise<any[]> {
   await initVectorDB();
   if (!client) throw new Error("Qdrant not initialized");
 
   try {
     const queryVector = await embedText(query);
-
     const results = await client.search(COLLECTION_NAME, {
       vector: queryVector,
       limit: topK,
       with_payload: true,
-      filter: {
-        must: [
-          {
-            key: "folderId",
-            match: {
-              value: folderId
-            }
-          }
-        ]
-      }
     });
 
-    // Map to match LanceDB format expected by other tools
-    return results.map(r => ({
+    return results.map((r) => ({
       id: r.id,
-      vector: r.vector,
-      ...r.payload
+      ...r.payload,
     }));
   } catch (err: any) {
     console.error("Qdrant search error:", err.message);
@@ -133,58 +307,74 @@ export async function searchProjectMemory(folderId: string, query: string, topK:
   }
 }
 
-export async function deleteProjectDocument(folderId: string, fileId: string): Promise<void> {
+/**
+ * Exhaustive full-text search using Qdrant's inverted index on the `text` field.
+ * Uses whitespace tokenizer → API paths like /v1/foo/bar match as single tokens.
+ * Paginates through all results server-side (no full collection scan in JS).
+ */
+export async function exactSearchChunks(
+  term: string,
+  limit: number = 50
+): Promise<any[]> {
   await initVectorDB();
-  if (!client) return;
+  if (!client) throw new Error("Qdrant not initialized");
 
-  await client.delete(COLLECTION_NAME, {
-    filter: {
-      must: [
-        { key: "folderId", match: { value: folderId } },
-        { key: "file_id", match: { value: fileId } }
-      ]
-    }
-  });
-  console.error(`Deleted old chunks from Qdrant for ${folderId} / ${fileId}`);
-}
+  const filter = {
+    must: [
+      {
+        key: "text",
+        match: { text: term.toLowerCase() },
+      },
+    ],
+  };
 
-export async function checkProjectDocumentExists(folderId: string, fileId: string): Promise<boolean> {
-  await initVectorDB();
-  if (!client) return false;
-
-  const res = await client.count(COLLECTION_NAME, {
-    filter: {
-      must: [
-        { key: "folderId", match: { value: folderId } },
-        { key: "file_id", match: { value: fileId } }
-      ]
+  const results: any[] = [];
+  let offset: string | number | null | undefined = undefined;
+
+  // Paginate until all matching points are collected or limit is reached
+  do {
+    const page: { points: any[]; next_page_offset?: string | number | null } =
+      await (client as any).scroll(COLLECTION_NAME, {
+      filter,
+      with_payload: true,
+      with_vector: false,
+      limit: Math.min(100, limit - results.length),
+      ...(offset !== undefined ? { offset } : {}),
+    });
+
+    for (const point of page.points) {
+      results.push({ id: point.id, ...point.payload });
     }
-  });
-  return res.count > 0;
+    offset = page.next_page_offset;
+  } while (offset != null && results.length < limit);
+
+  return results;
 }
 
-export async function getProjectDocumentMetadata(folderId: string): Promise<Record<string, string>> {
+/**
+ * Upsert agent note với random UUID (không có fileId).
+ */
+export async function upsertAgentNote(text: string): Promise<void> {
   await initVectorDB();
-  if (!client) return {};
-
-  const res = await client.scroll(COLLECTION_NAME, {
-    filter: {
-      must: [
-        { key: "folderId", match: { value: folderId } },
-        { key: "source", match: { value: "google_drive" } }
-      ]
-    },
-    limit: 10000,
-    with_payload: ["file_id", "modified_time"],
-    with_vector: false
-  });
-
-  const fileMap: Record<string, string> = {};
-  for (const r of res.points) {
-    if (r.payload && r.payload.file_id && r.payload.modified_time) {
-      fileMap[r.payload.file_id as string] = r.payload.modified_time as string;
-    }
-  }
+  if (!client) throw new Error("Qdrant not initialized");
 
-  return fileMap;
+  const vector = await embedText(text);
+  await client.upsert(COLLECTION_NAME, {
+    wait: true,
+    points: [
+      {
+        id: uuidv4(),
+        vector,
+        payload: {
+          text,
+          title: "Agent Note",
+          block_index: 0,
+          block_hash: "",
+          source: "agent",
+          offset: 0,
+        },
+      },
+    ],
+  });
+  console.error("Upserted agent note to Qdrant.");
 }