Tuning¶
The defaults are conservative for a single-user machine. These are the main knobs to revisit as a collection grows. All are environment variables — see the Configuration overview and the full Settings reference.
Ranking, chunking, jobs, and clustering¶
| Setting | Default | Purpose |
|---|---|---|
REFINDERY_CHUNKING__TARGET_TOKENS |
448 |
Desired chunk size. |
REFINDERY_CHUNKING__OVERLAP_TOKENS |
64 |
Context repeated across chunks. |
REFINDERY_CHUNKING__HARD_MAX_TOKENS |
512 |
Maximum canonical chunk size. |
REFINDERY_FETCH__TIMEOUT_S |
10.0 |
Outbound fetch timeout. |
REFINDERY_FETCH__MAX_BYTES |
10000000 |
Maximum fetched response size. |
REFINDERY_JOBS__MAX_ATTEMPTS |
5 |
Attempts before a job becomes dead. |
REFINDERY_JOBS__LEASE_MINUTES |
15 |
Recovery lease for in-flight work; handlers are cancelled at expiry. |
REFINDERY_JOBS__HANDLER_TIMEOUT_S |
(lease) | Override the handler cancellation timeout. |
REFINDERY_CLUSTER__MIN_PAGES |
50 |
Pages required for the first cluster run. |
REFINDERY_CLUSTER__MIN_NEW_PAGES |
20 |
New pages required for an idle-triggered run. |
REFINDERY_SEARCH__RECENCY_HALF_LIFE_DAYS |
(unset) | Optional ranking decay toward recent pages. |
Chunking is model-independent
Chunk sizes are canonical and shared across all embedding models, so all models embed the same spans. Changing them re-chunks the corpus. The hard max must not exceed the smallest registered model's token budget — see Embedding models.
Reranking¶
Set the reranker to none for fusion-only search with no reranking provider
call:
Otherwise choose an API reranker (Cohere, Voyage) or a local cross-encoder via
REFINDERY_RERANKER__*. See Searching.
Provider resilience¶
External provider calls (embedding, reranking, LLM) run behind a per-provider circuit breaker with in-call retry — see Operations:
| Setting | Default | Purpose |
|---|---|---|
REFINDERY_RESILIENCE__BREAKER_FAILURE_THRESHOLD |
5 |
Consecutive failures before a breaker opens. |
REFINDERY_RESILIENCE__BREAKER_COOLDOWN_S |
30.0 |
Fast-fail window before a probe is admitted. |
REFINDERY_RESILIENCE__RETRY_ATTEMPTS |
3 |
In-call attempts per provider call. |
REFINDERY_RESILIENCE__RETRY_BASE_DELAY_S |
0.25 |
First retry backoff. |
REFINDERY_RESILIENCE__RETRY_MAX_DELAY_S |
2.0 |
Retry backoff cap. |
REFINDERY_RESILIENCE__EMBED_TIMEOUT_S |
60.0 |
Per-attempt embedding timeout. |
REFINDERY_RESILIENCE__RERANK_TIMEOUT_S |
15.0 |
Per-attempt rerank timeout. |
REFINDERY_LLM__TIMEOUT_S |
30.0 |
Per-attempt LLM completion timeout. |
Clustering schedule¶
Clustering runs on an idle trigger by default. To add a cron schedule, set a one-to-five-field crontab expression:
Algorithm, reducer, and sizing live under REFINDERY_CLUSTER__* — see
Clustering.
Related¶
- Searching — where these knobs take effect.
- Clustering — sizing and triggers.
- Operations — job lease behavior.