feat(embedder): add configurable batch_size

[ibreakshit]

Summary

Adds an optional embedder.batch_size config option that caps the number of inputs per embedding API request. Default is 2000 (unchanged).

Motivation

On slow or self-hosted OpenAI-compatible embedding endpoints, grepai's initial scan packs the entire batch (up to MaxBatchSize = 2000) into a single /v1/embeddings request. When the endpoint is slow, that one request exceeds the embedder's client timeout and the whole initial index fails:

initial indexing failed: failed to embed batches: failed to send request to OpenAI:
Post ".../v1/embeddings": context deadline exceeded (Client.Timeout exceeded while awaiting headers)

There is currently no config-level way to work around this. Lowering batch_size keeps each request short so indexing completes reliably as more, smaller requests (slower overall, but it finishes). On throughput-bound endpoints it pairs well with parallelism: 1.

embedder:
  provider: openai
  endpoint: http://my-slow-gateway:4000/v1
  model: qwen3-embedding-8b
  parallelism: 1
  batch_size: 20

Changes

• config: EmbedderConfig.BatchSize (yaml: batch_size)
• embedder: additive FormBatchesWithSize(files, maxBatchSize); FormBatches now delegates to it with the MaxBatchSize default (no behavior change). Values <= 0 or > MaxBatchSize fall back to the default.
• indexer: per-Indexer batchSize (defaults to MaxBatchSize) + SetBatchSize; the batched embed path uses FormBatchesWithSize.
• cli/watch: applies cfg.Embedder.BatchSize to the indexer for both the single-project and workspace paths.
• Tests (FormBatchesWithSize custom size, default fallback, FormBatches delegation) and docs.

Backward compatibility

Fully backward compatible. Unset / <= 0 / > MaxBatchSize all use the existing 2000 default; FormBatches behaves exactly as before. The MaxBatchTokens limit still applies.

Testing

go build ./cmd/grepai
go test ./embedder/... ./config/... ./indexer/...   # all pass
gofmt -l ...                                         # clean

🤖 Generated with Claude Code