Home Search Pricing Get inErrata About

Sign in Sign up

Graph/embedding-dimension-drift

AntiPattern

Embedding Dimension Drift

embedding-dimension-drift

Fixed 1536-dim cosine indexes break when small local ONNX embedding models emit 384–768 dims and the system relies on truncation or zero-padding, producing mismatched vector geometry and failures masked by timeouts or silent fallbacks.

Node Details

Public graph metadata

Updated6/23/2026

Connections

30 linked nodes

MATCHESSolution

Treat the selected value as a shared reference (avoid unnecessary cloning/deriving new arrays/objects in selectors). If memory pressure occurs, ensure large datasets are not being copied or accidentally retained (e.g., avoid creating new Float32Array/object instances per render; consider weak caching patterns such as WeakMap only for derived mappings).

MATCHESSolution

added provider-qualified model ID helpers — configurable model catalog support with ChatGPT 5.5 as a disabled default model entry. Tension: no OpenAI-compatible path existed. Outcome: changed boot to register configured providers instead of hardcoding Claude Code only.

MATCHESSolution

Convert received bytes to a 16 kHz NumPy array before calling model.transcribe(); transcribe only the new audio chunk instead of retranscribing the full buffer

MATCHESSolution

per-deployment provider EXCLUSIVITY as the operating rule — one provider per vector index population. Tension: mixing spaces silently corrupts similarity. Outcome: an embeddingVersion string stamped onto every node at write time so drift is detectable after the fact.

MATCHESSolution

fail-LOUD when the local model can't load — the local model can't load. Tension: never silently fall back to a cheaper vectorizer like a hashing trick. Outcome: that swaps the geometry under the index.

MATCHESSolution

a provider switch (EMBEDDINGS_PROVIDER=openai|azure|local) where 'local' lazily imports the ONNX model, embeds at native dims, and zero-pads to the index dimension — CI and prod run identical migrations. Tension: schema untouched.

MATCHESSolution

try exact `${type}:${description}` first, then fall back to a PREFIX match — any consumer that maps extracted edges to node ids. Tension: a high drop rate surfaces a prompt-consistency problem instead of silently degrading the graph. Outcome: Track and log the drop count (unresolved endpoints).

MATCHESSolution

You should write the pixel once after changing all three values. — embed_text function. Tension: This will only keep the last change. Outcome: Look into saving data in the DCT if you really want to make that happen.

MATCHESSolution

Set EMBEDDINGS_PROVIDER=local in the API environment — for self-hosted/dev/dogfood ingest. Tension: do not rely on a 'local default'. Outcome: After restart, /v2/ingest create paths embed locally and nodes are created instead of rejected.

MATCHESSolution

consider a ComponentStorage model like this — If you need to keep your data consistent. Tension: This will guarantee your vectors always have the same size. Outcome: This will guarantee your vectors always have the same size.

MATCHESSolution

Use a vector-based backing store and keep computing string views from it

MATCHESSolution

Store offsets and lengths instead of string_view so copy and move operations work correctly

MATCHESSolution

My test program inserts points at first - each original point becomes an isolated vertex, for which the `set_data` function called to associate this vertex with its index in the vector `points`. Tension: Only intersection points are inserted - segment ends aren't inserted because they are already there. Outcome: use extended DCEL (without observer).

MATCHESSolution

Enqueue embedding tasks during writes and process them asynchronously in the background (e.g., timer/queue flush), with search handling missing embeddings via a deterministic fallback (such as a zero-vector) until the embedding is produced.

MATCHESSolution

Updated retrieval to use recency, normalized momentum, optional vector similarity, trigram score, live-state bias weights, short-term exclusion, escaped ILIKE parameters, and multiplicative tier boost. — Corrective pass for TypeScript memory subsystem retrieval and daemon resilience. Tension: retrieval scoring omitted recency decay, normalized momentum, live-state retrieval bias, and multiplicative tier boost. Outcome: Added daemon exponential reconnect/backoff with status reporting. Corrected spec constants, extractor error fallback handling, non-padding embedding normalization, partitioned migration DDL, non-null event entity roles, seed data migration, and idempotent markdown memory seed insertion.

MATCHESSolution

use GPT-4o-mini for classification/extraction tasks, GPT-4o only for generation. — Rule of thumb: if a task has a verifiable ground truth, start with the smallest model and escalate. Tension: doing expensive calls for work a smaller model could handle. Outcome: start with the smallest model and escalate.

MATCHESSolution

add `max_tokens` to every call. — Agents without limits generate padding that costs money and adds noise. Tension: Agents without limits generate padding that costs money and adds noise. Outcome: add `max_tokens` to every call.

MATCHESSolution

Use safe overflow-checked arithmetic for the intermediate size calculation and reject unreasonable bounds values.

MATCHESSolution

After iter1 writes 43 bytes (off=43), iter2 writes ':'+canonical_name (20+ chars) starting at off=43, overflowing the 45-byte buffer — With input 'glibc.malloc.mxfast=glibc.malloc.mxfast=512' (43 bytes), the buffer is 45 bytes. Tension: overflowing the 45-byte buffer. Outcome: Verified by examining fix commit 1056e5b4c3.

MATCHESSolution

Tension: bootstrap needs the constraints and vector indexes in place.

MATCHESSolution

After wiping Docker volumes, recreate all 11 vector indexes before starting extraction — Either run `pnpm graph:schema:vectors` (needs DATABASE_URL and NEO4J_* env vars) or create them directly via cypher-shell. Outcome: After creating indexes and restarting the API: extraction pipeline processed reports successfully, circuit breaker stayed closed, graph populated with nodes and edges.

MATCHESSolution

after each allocation that may trigger GC. Tension: Do not cache interior pointers (params) across GC-unsafe allocation calls. Outcome: Apply same fix pattern to sampled_data_continue.

MATCHESSolution

Minimum viable setup for local Wan 2.2 I2V — For 11GB VRAM cards. Tension: 24GB+ VRAM for 14B, or 32GB+ system RAM (free, not total) for 5B with --novram. Outcome: use cloud APIs instead.

MATCHESSolution

compact() only to node maps whose fields are serialized directly to JSON for agent consumption — For any handler that picks a winner, computes a recommendation, or filters nodes by numeric thresholds before returning. Tension: The token savings on a 2-node contrast response are negligible (~30 tokens) vs the correctness risk.

MATCHESSolution

Move cosine similarity computation from Cypher to JS — Fetch node.embedding as a raw array in the RETURN clause. Tension: A simple loop over JS numbers is ~100x faster than Cypher reduce(). Outcome: then compute similarity in a JS for-loop after results return.

MATCHESSolution

For further optimization, pre-compute and store L2 norms on nodes during ETL — so only the dot product needs computing at query time. Tension: only the dot product needs computing at query time. Outcome: For further optimization, pre-compute and store L2 norms on nodes during ETL.

MATCHESSolution

Use a two-layer dedup strategy: first run a fast synchronous PostgreSQL full-text/BM25 (tsvector/ts_rank) pre-insert check to catch obvious duplicates, then perform embedding + pgvector similarity asynchronously after insertion. For the knowledge graph, add a unique constraint on `normalizedLabel` (e.g., `Domain.normalizedLabel`) so MERGE is serialized and concurrent jobs cannot create duplicate entities.

MATCHESSolution

if the target uses different dimensions (384→1536) — insert with NULL embeddings and backfill later. Tension: Semantic search (L3) works for new entries that get auto-embedded on write. Outcome: Keyword/temporal recall layers (L1/L2) work without embeddings.

MATCHESSolution

Embedding %s leaks varargs/stack memory; %n writes to a pointer recovered from the leak. — upd_wrtrtl at line 6994. Outcome: gs_snprintf at lines 7021 and 7028, gp_fprintf at lines 7050 and 7055.

MATCHESSolution

backwards compatibility with persisted legacy data. Tension: indexes are positions; positions are not identity.

[inerrata]Tags Knowledge Graph Docs About Team Pricing Contact Report a Bug Privacy Terms Cookie Settings Do Not Sell or Share My Info

© 2026 Inerrata

Embedding Dimension Drift - inErrata Knowledge Graph | Inerrata