[Leo310]

Fixes #485.

What was wrong

Ollama embedding runs stalled 10–30 minutes per file with the GPU idle. Two independent causes:

1. HTTP 400 retried six times. LangChain's default retry handler skips 4xx, but it reads the status from status / statusCode / response.status. The Ollama client's ResponseError exposes it as status_code, so "the input length exceeds the context length" was treated as transient. Reproduced against a mock server that always answers 400: 7 requests, 84.5 s for one embed call. The bulk indexer's per-entry fallback then repeats that for every chunk in the failed batch, which is where the multi-minute stalls come from.
2. Chunks sized to the wrong limit. The budget came from /api/show's architecture context (40960 tokens for qwen3-embedding) × 4 chars/token, but Ollama's server enforces min(num_ctx, model context), and its num_ctx default is VRAM-chosen (4k+) and not readable from the client. Every long note overshot by 5×.

What changed

• createTransportedOllamaEmbeddings installs an onFailedAttempt that rethrows on 4xx status_code and on cancellation, leaving the default retry budget for transport failures and 5xx.
• The Ollama template pins num_ctx on embed requests to OLLAMA_EMBED_NUM_CTX (8192), and hydrateEmbeddingModel caps Ollama input budgets to the same number, so client and server agree on the limit.
• truncate: true on Ollama embeddings as a backstop: a truncated vector still finds the note by its opening, a rejected one drops it from the index.
• CHARS_PER_TOKEN 4 → 3 so dense content (JSON, code, tables, non-Latin) stops overshooting in general. Affects chunk sizing only on (re)index; existing indexes are untouched.

Tests

• test/providers/ollamaEmbeddings.test.ts: handler rules, and an end-to-end mock-400 that asserts exactly one request (fails by timeout before the fix), plus the template's num_ctx/truncate wiring.
• test/lib/modelMetadataNormalizer.test.ts: the Ollama cap, smaller contexts left alone, other providers uncapped.

bun run check, format, lint, test (1682 passing) and a dev build are clean.

[greptile-apps]

Greptile Summary

This PR aligns Ollama embedding chunk budgets with the num_ctx sent to the server and prevents permanent 4xx responses or cancellations from being retried. It also makes bulk indexing preserve chunk-0 completion semantics when any chunk from a note fails.

• Pins Ollama embedding requests and metadata budgets to an 8192-token context.
• Uses a more conservative character-to-token estimate and enables truncation as a fallback.
• Stops retries for Ollama 4xx responses and canceled requests while retaining retries for transient failures.
• Prevents failed notes from receiving the chunk-0 completion marker.
• Adds regression coverage for request behavior, metadata normalization, and Ollama configuration.

Confidence Score: 5/5

The PR appears safe to merge; no outstanding correctness, security, or repository-rule issues remain.

The prior partial-index concern was manually resolved after the current failure tracking preserved chunk-0 as a completion marker, and the present code consistently aligns Ollama request limits, chunk sizing, retry behavior, and recovery semantics.

Important Files Changed

Filename	Overview
src/providers/chatProviders.ts	Adds an Ollama-specific failed-attempt handler that immediately stops retries for permanent client errors and cancellation.
src/providers/ollama.ts	Pins embedding request context and enables truncation as a final safeguard against oversized inputs.
src/lib/modelMetadataNormalizer.ts	Caps Ollama's advertised embedding input budget to the context configured on requests.
src/providers/ollamaModels.ts	Defines the shared Ollama embedding context constant used by request construction and chunk budgeting.
src/vectorstore/VectorStoreService.ts	Uses more conservative chunk sizing and prevents a failed note from receiving its chunk-0 completion marker.
test/providers/ollamaEmbeddings.test.ts	Covers retry classification, single-request behavior for HTTP 400, and Ollama request options.
test/lib/modelMetadataNormalizer.test.ts	Covers Ollama context capping, smaller model contexts, defaults, and unaffected non-Ollama providers.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Read Ollama model metadata] --> B[Cap input budget at 8192 tokens]
  B --> C[Split note using conservative token estimate]
  C --> D[Send embedding request with num_ctx 8192]
  D --> E{Response}
  E -->|Success| F[Write chunks with chunk 0 last]
  E -->|4xx or cancellation| G[Stop retrying]
  E -->|Transport failure or 5xx| H[Use retry budget]
  G --> I[Mark note failed and suppress remaining chunks]
  I --> J[Leave chunk 0 absent for later reindex]

Loading

_{Reviews (3): Last reviewed commit: "Merge branch 'main' into fix/ollama-embe..." | Re-trigger Greptile}