002: Redis Caching

Date: 2026-05-15

Context

Queries requiring LLM planning are latency-intensive. Many questions are repeatedly asked. Caching is essential, but simply caching API responses is dangerous in finance, as datasets update and cached values become instantly stale.

Decision

Implement a Redis-based cache accompanied by a dataset-version-aware invalidation strategy.

Consequences

• Positive: Frequent queries are answered in milliseconds. If the underlying dataset changes (e.g., a corporate action is applied retroactively), we can increment the dataset version and intelligently purge or ignore stale cache lines.
• Negative: Requires maintaining a CacheLineage mapping in the database to track which cache keys belong to which dataset versions.

Alternatives Considered

• In-memory LRU Cache: Does not scale horizontally across multiple API instances.
• TTL-based invalidation only: Rejected as it allows serving incorrect financial data for up to the TTL duration after an update.