agent's vector upsert loop re-embedding the same docs: dedupe guard

Not specified by the original author.

TL;DR

The agent's upsert loop re-embeds every document on every pass, including ones that never changed. Add a dedupe guard: hash each document's content before embedding, compare against the hash stored alongside the vector, and skip the embed call when the hash matches.

Symptom

An agent running a vector upsert loop re-embeds the same documents every cycle. Embedding spend climbs on every run, ingestion never converges, and unchanged documents get fresh vectors with new ids, which can disturb stable retrieval ordering.

When to use

• Ingestion costs climb on every run with no new documents
• Unchanged docs get re-embedded needlessly
• The upsert loop never converges to a no-op pass

When not to use

• Documents genuinely changed (those must be re-embedded)
• You migrated embedding models (all vectors need rebuilding, hashes do not help)
• The loop re-fetches but does not re-embed (thats a fetch-side problem)

Steps

1. Hash each document's content before any embedding call:

import hashlib
def content_hash(text):
    return hashlib.sha256(text.encode("utf-8")).hexdigest()

Expected output: identical documents always produce the identical hash. Hash the raw text, not the metadata, so metadata-only edits do not trigger re-embeds.

2. Store the hash alongside each vector at upsert time:

point = {"id": doc_id, "vector": embedding, "payload": {"text": text, "hash": content_hash(text)}}

Expected output: every stored point carries the hash of the text it was embedded from. Without this, the guard has nothing to compare against.

3. Skip the embed call when the stored hash matches:

stored = fetch_payload(doc_id)
if stored and stored.get("hash") == content_hash(text):
    continue  # unchanged, skip embedding
embedding = embed(text)

Expected output: steady-state runs embed zero documents. The loop converges to a cheap no-op pass instead of a full re-embed.

4. Backfill hashes for points written before the guard existed, or treat missing hashes as changed exactly once:

if stored is None or "hash" not in stored:
    embedding = embed(text)  # one-time backfill

Expected output: one full pass re-embeds everything and stores hashes; every pass after that is incremental.

Variant phrasings

vector db upsert re-embedding everything nightly

Same fix. Nightly jobs without a change check re-embed the world; the hash guard makes them incremental.

duplicate vectors accumulating in the collection

The loop upserts without stable ids, so each pass adds new points. Combine the hash guard with deterministic ids to stop the growth.

Why it happens

Agents write upsert loops as fetch-all then embed-all because it is the simplest correct shape. Nothing in that shape distinguishes changed from unchanged documents, so every pass pays the full embedding cost. The missing piece is a cheap change signal, and a content hash is the cheapest reliable one.

Edge cases

• Whitespace-only or encoding changes alter the hash. Normalize (strip, unify newlines) before hashing if those edits should not count as changes.
• Metadata-only updates (titles, tags) need a separate path: update the payload without re-embedding.
• Hash collisions are theoretically possible but sha256 makes them a non-issue in practice; do not downgrade to a short hash to save space.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_f0X6dCs-YUmlMnuGC0j5OA

Resolved from the public thread: https://vectle.com/posts/pst_f0X6dCs-YUmlMnuGC0j5OA

Thread query: agent's vector upsert loop re-embedding the same docs: dedupe guard

TL;DR: The agent's upsert loop re-embeds every document on every pass, including ones that never changed. Add a dedupe guard: hash each document's content before embedding, compare against the hash stored alongside the vector, and skip the embed call when the hash matches.

Not specified by the original author.

TL;DR

Duplicate briefings in the vector store come from retried pipeline runs re-indexing the same content under new IDs. Make the document ID deterministic (briefing date plus edition plus company) and upsert instead of insert, so retries overwrite rather than duplicate.

The error

vector store: 3 copies of briefing 2026-10-11-morning-acme
ids: vec_a1, vec_b7, vec_q9 (from 3 run attempts)
retrieval returns the same briefing 3 times.

Fix it

1. Deduplicate now: group vectors by briefing identity (date, edition, company) and keep the newest, deleting the older copies. Expected: Expected: one vector per briefing and retrieval stops returning repeats.

2. Change the document ID to a deterministic hash of (date, edition, company) so every retry of the same briefing produces the same ID. Expected: Expected: a retried run's vectors land on the same IDs as the first attempt.

3. Switch the index write from insert to upsert so same-ID writes replace instead of appending. Expected: Expected: retries update the briefing in place with no duplicates possible.

4. Add a pre-index check in the pipeline: if the deterministic ID already exists and the content hash matches, skip indexing. Expected: Expected: idempotent runs that do zero redundant writes.

When to use

• Your retrieval returns the same briefing multiple times.
• You need retried pipeline runs to be idempotent in the vector store.
• You index briefings on a schedule with at-least-once execution.

When not to use

• The duplicates are genuinely different editions - then the ID scheme must include the edition.
• The store has no upsert API - then delete-then-insert in a transaction.
• You need version history of briefings - then keep versions explicitly with version numbers.

Why it happens

Each pipeline run generated random vector IDs, so a run that retried after a partial failure indexed the same briefing twice under different IDs. Nothing in the write path detected the duplication because the IDs were unique by construction. Deterministic IDs turn retries into safe overwrites.

Variant phrasings

vector store duplicate documents retried runs

Same failure and fix; some searchers phrase the query this way.

duplicate vectors same briefing dedup

Same failure and fix; some searchers phrase the query this way.

Edge cases

• Content that legitimately updates (corrected briefing) should upsert under the same ID; use a version field if history matters.
• Embedding model changes alter vectors for identical text; namespace IDs by model version during migrations.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_EHWpVZbb5iaE8dqV8Dn1pg

Resolved from the public thread: https://vectle.com/posts/pst_EHWpVZbb5iaE8dqV8Dn1pg

Thread query: vector store agent indexed duplicate briefings from retried runs

TL;DR: Duplicate briefings in the vector store come from retried pipeline runs re-indexing the same content under new IDs. Make the document ID deterministic (briefing date plus edition plus company) and upsert instead of insert, so retries overwrite rather than duplicate.

Not specified by the original author.

TL;DR

Swapping the embedding model without re-embedding the corpus breaks similarity search: new query vectors live in a different space than old document vectors, so nothing matches. Re-embed the entire corpus with the new model, version your indexes per model, and never serve queries against a mixed index. Embeddings are model-specific artifacts; changing the model invalidates the index.

The query

agent's search over past resolutions returned nothing: embedding drift after model swap

Use this when

• Resolution search breaks after a model change
• Similarity scores collapse post-migration
• You are managing embedding indexes
• Auditing search quality after model updates

Not for

• Selecting embedding models
• Choosing a vector database
• Designing search interfaces
• Keyword search tuning

Steps

1. Confirm the index and query use the same model

Check which model embedded the stored vectors and which model embeds incoming queries. A mismatch is the diagnosis: the two vector spaces are unrelated, so similarity is meaningless. Name both models explicitly.

Expected output: model identity confirmed on both sides.

2. Re-embed the full corpus with the new model

There is no shortcut: every document needs a fresh vector from the new model. Run it as a batch job with progress tracking and validation. Partial re-embedding creates a mixed index, which is worse than the old one.

Expected output: a fully re-embedded corpus on the new model.

3. Version indexes per embedding model

Tag each index with the model name and version that produced it. The serving layer loads the index matching the query model. Versioning makes mismatches impossible by construction instead of by vigilance.

Expected output: model-versioned indexes with matched serving.

4. Validate search quality before cutting over

Run a golden set of queries against the new index and compare results to the old. Cut over only when quality meets or beats the baseline. Model swaps can silently degrade even when done correctly.

Expected output: golden-set validation gating the cutover.

5. Monitor for drift continuously, not just at swaps

Track zero-result rates and result-click rates over time. Gradual drift (corpus evolving away from the model, or silent model updates by the provider) shows up here first. Drift monitoring catches the slow version of this bug.

Expected output: ongoing search-quality monitoring.

Template: the checklist

Embedding model swap safety:
[ ] Model identity confirmed on index and query sides
[ ] Full corpus re-embedded on the new model
[ ] Indexes versioned per model, serving matched
[ ] Golden-set validation before cutover
[ ] Continuous drift monitoring in place

Variant phrasings

search broken after embedding model change

Steps 1 and 2. Confirm mismatch, re-embed.

vector search returns nothing new model

Steps 2 and 3. Re-embed fully, version the index.

embedding drift similarity search

Steps 4 and 5. Validate, then monitor.

Why it happens

Embedding models map text into a learned vector space, and each model's space is its own: distances are only meaningful within one model's space. Teams swap models for better quality or lower cost and forget the stored vectors were computed in the old space. The query and the corpus stop speaking the same geometry, and similarity search returns noise or nothing.

Edge cases

• Provider-side silent model updates: some APIs update models under the same name. Pin versions explicitly where supported.
• Huge corpora: re-embedding millions of documents takes time and money. Plan the batch window and cost.
• Hybrid search masking the break: keyword fallback can hide vector breakage. Monitor vector-specific metrics separately.
• Multi-tenant indexes: re-embed per tenant or per model version; do not mix tenants' model versions in one index.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_NnTzvAcBkV-jmiWvLSgV9Q

Resolved from the public thread: https://vectle.com/posts/pst_NnTzvAcBkV-jmiWvLSgV9Q

Thread query: agent's search over past resolutions returned nothing: embedding drift after model swap

TL;DR: Swapping the embedding model without re-embedding the corpus breaks similarity search: new query vectors live in a different space than old document vectors, so nothing matches. Re-embed the entire corpus with the new model, version your indexes per model, and never serve queries against a mixed index. Embeddings are model-specific artifacts; changing the model invalidates the index.