Truncate embedding input by tokens at the model's window, not by a character cap

Ouverte
#28 0 commentaires 0 réactions 1 personne assignée Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

@kriszyp y travaille déjà.

Depuis le 28/10/2026.

Évaluation

Cette issue n'a pas encore été évaluée.

Description

Problem

limits.embeddingInputChars (added in 0.1.2, branch fix/bound-embedding-input) cuts every embedding input to a character count, default 6,000. Characters are a proxy for tokens, and it misses in both directions:

  • Too long for token-dense text. 6,000 characters of CJK text (about 1 token per character) or code (about 3 characters per token), plus any task prefix a host backend adds, can still exceed a 2,048-token window. The backend answers HTTP 400, and the write fails with a 500 as before.
  • Too short for large windows. A model that holds 8k tokens embedded a 20,000-character memory whole before 0.1.2. With the default it now embeds only the first 6,000 characters, so vector recall on the tail is lost unless the operator raises the cap. Full-text search still covers the tail.

Proposal

Truncate by tokens at the model's window, in Harper core. Add an EmbedOpts truncation option (for example truncate: 'end'), which the openai backend forwards as vLLM's truncate_prompt_tokens and other backends honor as they can. Today the openai backend builds { model, input } only, so a plugin cannot ask for this. Once core supports it, the plugin passes the option and keeps the character cap only as a backstop for backends without truncation, with a default that no longer affects large-window models.

Related: Re-embed memories when embeddingModel changes vector size. Changing the bound does not re-embed existing rows either.

Review-Deferral: 1913aa1ba8c0ab82 severity=minor category=correctness

🤖 Generated by Claude Opus 5.5 (Anthropic, Claude Code dev-agent); posted via @kriszyp.

Langage dominant
JavaScript
Étoiles
0
Forks
0
Merge moyen
20 h 29 min
PR mergées (30 j)
28

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de HarperFast/agent-memory

Toutes les issues de HarperFast/agent-memory

Issues similaires

Plus d'issues JavaScript

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.