@vellumai/assistant 0.5.16 → 0.6.1

package/src/memory/journal-memory.ts

@@ -1,224 +0,0 @@
-import { readdirSync, readFileSync, statSync } from "node:fs";
-import { join } from "node:path";
-
-import { and, eq } from "drizzle-orm";
-import { v4 as uuid } from "uuid";
-
-import { getLogger } from "../util/logger.js";
-import { getWorkspaceDir } from "../util/platform.js";
-import { type DrizzleDb, getDb } from "./db.js";
-import { computeMemoryFingerprint } from "./fingerprint.js";
-import { enqueueMemoryJob } from "./jobs-store.js";
-import { memoryItems, memoryItemSources } from "./schema.js";
-
-const log = getLogger("memory-journal");
-
-/**
- * Process a single journal `.md` file: read content, derive subject, compute
- * fingerprint, upsert to DB, and enqueue an embed job.
- *
- * Returns `true` if a new memory item was inserted, `false` if it already
- * existed (or was skipped).
- */
-function upsertSingleJournalFile(
-  filepath: string,
-  filename: string,
-  messageCreatedAt: number,
-  scopeId: string,
-  messageId: string,
-  db: DrizzleDb,
-): boolean {
-  const content = readFileSync(filepath, "utf-8");
-
-  // Derive subject from filename:
-  // strip .md extension, strip leading date prefix, replace hyphens with spaces, capitalize first letter
-  const basename = filename.replace(/\.md$/, "");
-  const withoutDate = basename.replace(/^\d{4}-\d{2}-\d{2}-?/, "");
-  const withSpaces = withoutDate.replace(/-/g, " ");
-  const subject =
-    withSpaces.length > 0
-      ? withSpaces.charAt(0).toUpperCase() + withSpaces.slice(1)
-      : basename;
-
-  const fingerprint = computeMemoryFingerprint(
-    scopeId,
-    "journal",
-    subject,
-    content,
-  );
-
-  const existing = db
-    .select()
-    .from(memoryItems)
-    .where(
-      and(
-        eq(memoryItems.fingerprint, fingerprint),
-        eq(memoryItems.scopeId, scopeId),
-      ),
-    )
-    .get();
-
-  let memoryItemId: string;
-  let inserted = false;
-
-  if (existing) {
-    memoryItemId = existing.id;
-    db.update(memoryItems)
-      .set({
-        lastSeenAt: messageCreatedAt,
-        status: "active",
-      })
-      .where(eq(memoryItems.id, existing.id))
-      .run();
-  } else {
-    memoryItemId = uuid();
-    db.insert(memoryItems)
-      .values({
-        id: memoryItemId,
-        kind: "journal",
-        subject,
-        statement: content,
-        status: "active",
-        confidence: 0.95,
-        importance: 0.8,
-        fingerprint,
-        sourceType: "extraction",
-        sourceMessageRole: "assistant",
-        verificationState: "assistant_inferred",
-        scopeId,
-        firstSeenAt: messageCreatedAt,
-        lastSeenAt: messageCreatedAt,
-        lastUsedAt: null,
-        supersedes: null,
-        overrideConfidence: null,
-      })
-      .run();
-    inserted = true;
-  }
-
-  db.insert(memoryItemSources)
-    .values({
-      memoryItemId,
-      messageId,
-      evidence: content,
-      createdAt: Date.now(),
-    })
-    .onConflictDoNothing()
-    .run();
-
-  enqueueMemoryJob("embed_item", { itemId: memoryItemId });
-
-  return inserted;
-}
-
-/**
- * Scan the journal directory for `.md` files created during (or after) the
- * given message timestamp and upsert them as journal memory items with the
- * raw, unedited file content as the `statement`.
- *
- * Also scans immediate subdirectories (e.g. per-user folders like
- * `journal/sidd/`) so that user-scoped journal entries are indexed alongside
- * root-level files.
- *
- * This bypasses the LLM extraction layer entirely — journal memories are
- * stored verbatim so they are never summarised or rewritten.
- *
- * Returns the number of newly inserted items.
- */
-export function upsertJournalMemoriesFromDisk(
-  messageCreatedAt: number,
-  scopeId: string,
-  messageId: string,
-): number {
-  try {
-    const journalDir = join(getWorkspaceDir(), "journal");
-
-    let files: string[];
-    try {
-      files = readdirSync(journalDir);
-    } catch {
-      // Directory doesn't exist — no journal entries
-      return 0;
-    }
-
-    // Filter for .md files, excluding readme.md (case-insensitive)
-    const mdFiles = files.filter(
-      (f) =>
-        f.endsWith(".md") &&
-        !f.startsWith(".") &&
-        f.toLowerCase() !== "readme.md",
-    );
-
-    let upserted = 0;
-    const db = getDb();
-
-    for (const filename of mdFiles) {
-      try {
-        const filepath = join(journalDir, filename);
-        const stat = statSync(filepath);
-        if (!stat.isFile()) continue;
-
-        // Fall back to mtimeMs when birthtimeMs is unavailable (returns 0 on Linux ext4, NFS, Docker overlayfs)
-        const fileBirthtimeMs = stat.birthtimeMs > 0 ? stat.birthtimeMs : stat.mtimeMs;
-        // Only process files created during or after this message
-        if (fileBirthtimeMs < messageCreatedAt) continue;
-
-        if (upsertSingleJournalFile(filepath, filename, messageCreatedAt, scopeId, messageId, db)) {
-          upserted += 1;
-        }
-      } catch (err) {
-        log.warn(
-          { filename, err: err instanceof Error ? err.message : String(err) },
-          "Failed to process journal file for memory — skipping",
-        );
-      }
-    }
-
-    // Scan per-user journal subdirectories
-    for (const entry of files) {
-      try {
-        const subdirPath = join(journalDir, entry);
-        if (!statSync(subdirPath).isDirectory()) continue;
-
-        const subFiles = readdirSync(subdirPath).filter(
-          (f) =>
-            f.endsWith(".md") &&
-            !f.startsWith(".") &&
-            f.toLowerCase() !== "readme.md",
-        );
-
-        for (const filename of subFiles) {
-          try {
-            const filepath = join(subdirPath, filename);
-            const stat = statSync(filepath);
-            if (!stat.isFile()) continue;
-
-            // Fall back to mtimeMs when birthtimeMs is unavailable (returns 0 on Linux ext4, NFS, Docker overlayfs)
-            const fileBirthtimeMs = stat.birthtimeMs > 0 ? stat.birthtimeMs : stat.mtimeMs;
-            // Only process files created during or after this message
-            if (fileBirthtimeMs < messageCreatedAt) continue;
-
-            if (upsertSingleJournalFile(filepath, filename, messageCreatedAt, scopeId, messageId, db)) {
-              upserted += 1;
-            }
-          } catch (err) {
-            log.warn(
-              { filename, err: err instanceof Error ? err.message : String(err) },
-              "Failed to process journal file for memory — skipping",
-            );
-          }
-        }
-      } catch {
-        // Skip unreadable subdirectories
-      }
-    }
-
-    return upserted;
-  } catch (err) {
-    log.warn(
-      { err: err instanceof Error ? err.message : String(err) },
-      "Failed to scan journal directory for memories",
-    );
-    return 0;
-  }
-}

package/src/memory/query-builder.ts

@@ -1,47 +0,0 @@
-import { getSummaryFromContextMessage } from "../context/window-manager.js";
-import type { Message } from "../providers/types.js";
-
-const TRUNCATION_MARKER = "<truncated />";
-
-export interface MemoryQueryBuilderOptions {
-  maxUserRequestChars?: number;
-  maxSessionSummaryChars?: number;
-}
-
-/**
- * Build a deterministic memory recall query string from the current user
- * request plus any in-context session summary message.
- */
-export function buildMemoryQuery(
-  userRequest: string,
-  messages: Message[],
-  options?: MemoryQueryBuilderOptions,
-): string {
-  const maxUserRequestChars = options?.maxUserRequestChars ?? 2000;
-  const maxSessionSummaryChars = options?.maxSessionSummaryChars ?? 1200;
-
-  const requestText = clampSection(userRequest.trim(), maxUserRequestChars);
-  const sessionSummary = messages
-    .map((message) => getSummaryFromContextMessage(message))
-    .find((summary): summary is string => summary != null);
-
-  const content = requestText.length > 0 ? requestText : "(empty)";
-
-  if (sessionSummary && sessionSummary.trim().length > 0) {
-    const compactSummary = clampSection(
-      sessionSummary.trim(),
-      maxSessionSummaryChars,
-    );
-    return `${content}\n\nContext summary:\n${compactSummary}`;
-  }
-
-  return content;
-}
-
-function clampSection(value: string, maxChars: number): string {
-  if (maxChars <= 0) return "";
-  if (value.length <= maxChars) return value;
-  const half = Math.floor((maxChars - TRUNCATION_MARKER.length) / 2);
-  if (half <= 0) return value.slice(0, maxChars);
-  return value.slice(0, half) + TRUNCATION_MARKER + value.slice(-half);
-}

package/src/memory/query-expansion.ts

@@ -1,83 +0,0 @@
-/**
- * HyDE (Hypothetical Document Embeddings) query expansion for memory retrieval.
- *
- * Generates hypothetical memory documents that bridge the semantic gap between
- * how users query ("moments that changed everything") and how memories are
- * actually stored ("Sidd gave Velissa a collar on March 24"). The expanded
- * queries are embedded alongside the original query to improve recall.
- */
-
-import type { AssistantConfig } from "../config/types.js";
-import {
-  extractText,
-  getConfiguredProvider,
-  userMessage,
-} from "../providers/provider-send-message.js";
-import { getLogger } from "../util/logger.js";
-
-const log = getLogger("memory-query-expansion");
-
-const SYSTEM_PROMPT = `Generate 3 short hypothetical memory entries that would match this search query. Each should describe what a stored memory about this topic would contain — specific details, emotional context, relationship dynamics. Write from the perspective of stored memory items, not the query itself. Keep each under 80 words. Separate entries with ---`;
-
-/**
- * Generate hypothetical memory documents for a query using HyDE.
- *
- * Returns 1-3 hypothetical document strings that can be embedded alongside
- * the original query to improve semantic recall. Returns `[]` on any error
- * (provider unavailable, LLM failure, empty response) — the caller should
- * fall back to the raw query only.
- *
- * The raw query is NOT included in the returned array; the caller handles
- * that separately.
- */
-export async function expandQueryWithHyDE(
-  query: string,
-  _config: AssistantConfig,
-  signal?: AbortSignal,
-): Promise<string[]> {
-  try {
-    const provider = await getConfiguredProvider();
-    if (!provider) {
-      log.warn("No provider available for HyDE query expansion");
-      return [];
-    }
-
-    const response = await provider.sendMessage(
-      [userMessage(query)],
-      undefined,
-      SYSTEM_PROMPT,
-      {
-        config: {
-          modelIntent: "latency-optimized" as const,
-        },
-        signal,
-      },
-    );
-
-    const text = extractText(response);
-    if (!text) {
-      log.warn("Empty response from HyDE query expansion");
-      return [];
-    }
-
-    const entries = text
-      .split("---")
-      .map((entry) => entry.trim())
-      .filter((entry) => entry.length > 0)
-      .slice(0, 3);
-
-    if (entries.length === 0) {
-      log.warn("No entries parsed from HyDE query expansion response");
-      return [];
-    }
-
-    return entries;
-  } catch (err) {
-    if (err instanceof Error && (err.name === "AbortError" || err.name === "APIUserAbortError")) throw err;
-    log.warn(
-      { err: err instanceof Error ? err.message : String(err) },
-      "HyDE query expansion failed",
-    );
-    return [];
-  }
-}