quickstart.yaml

# CacheSeek quickstart configuration.
#
# Pair with:
#     from cacheseek import CacheService
#     cache = CacheService.from_config("quickstart.yaml")
#
# Defaults below run a local-file KV + FAISS vector setup with the
# encoder and reranker disabled — enough to wire the lifecycle end to
# end without GPU or external services. Flip the commented blocks below
# to plug real diffusion-side encoders (Qwen3-VL) and shared backends
# (Fluxon, Qdrant) into your serving environment.

# ─── Cache root ────────────────────────────────────────────────────
enable_latent_cache: true
cache_mode: read_write           # read_write | read_only | write_only
latent_cache_dir: ./cache
max_cache_size_gb: 10

# ─── KV backend ────────────────────────────────────────────────────
kv_store_type: local_file        # local_file | fluxon
# fluxon_config_path: /path/to/fluxon_config.yaml

# ─── Vector backend ────────────────────────────────────────────────
vector_store_type: faiss         # faiss | qdrant
faiss_index_dir: ./cache/faiss
# qdrant_url: http://127.0.0.1:6333
# qdrant_api_key: ~

vector_dim: 2048                 # match the embedding model output

# ─── Cache reuse strategy ──────────────────────────────────────────
cache_strategy_type: video_approximate
key_steps: [5, 10, 15, 20, 25]
max_skip_step: 5                 # upper bound on how many steps a hit may skip
lookup_mode: video
video_vector_collection: video
# Staircase skip-step: when rerank is on, skip depth is tiered by the rerank
# score (higher score → deeper skip), still capped by max_skip_step + saved_steps.
# Default table is the donor_drift ≤ 20% point from W05-0514-forced-K-sweep
# (docs/动态阈值实验.md §7). Has no effect here (rerank_enabled: false below).
# staircase_skip_enabled: false   # opt-in; flat max_skip_step skip by default
# skip_step_tau_table: {3: 0.63, 7: 0.85, 11: 0.85, 14: 1.01}

# ─── Encoder (off for the quickstart smoke; on for real serving) ───
video_embedding_enabled: false
# video_embedding_model_path: /path/to/Qwen3-VL-Embedding-2B
# video_embedding_device_id: 0
# text_embedding_model_path: /path/to/Qwen3-VL-Embedding-2B
# text_embedding_device_id: 0

# ─── Reranker (off for the quickstart smoke; on for tighter recall) ─
rerank_enabled: false
# rerank_model_path: /path/to/Qwen3-VL-Reranker-2B
# rerank_device_id: 0
# rerank_score_threshold: 0.80

# ─── Save path ─────────────────────────────────────────────────────
save_async_enabled: true
save_queue_size: 8
save_on_full: drop               # drop | sync | downgrade
flush_on_shutdown: true

# ─── Vector-update barrier ─────────────────────────────────────────
vector_wait_poll_s: 0.05
vector_wait_warn_s: 2.0
vector_wait_timeout_s: 120.0