FreshContext for RAG
Semantic relevance is not temporal relevance.
A RAG pipeline can retrieve the most semantically relevant document and still give the model context that is stale, weakly dated or no longer fit for the decision. FreshContext adds a source-aware judgment layer before generation.
- retriever
- semantic candidates
- FreshContext
- freshness + confidence + decision
- model
- decision-ready context
The failure mode
Cosine similarity does not know what changed yesterday.
A high semantic score answers “does this match?” It does not answer “is this still current enough to rely on?”
Stale but relevant
An older page can match the query perfectly while a newer policy, release or price has already replaced it.
One timer for every source
A single TTL treats fast-moving prices, news, API docs and stable definitions as if they decay at the same rate.
Unknown dates
When publication time is weak or missing, that uncertainty should travel with the context instead of disappearing before generation.
Fetch failures that look successful
Blocked, empty, malformed or rate-limited results should not silently enter the prompt as normal evidence.
Source-aware decay
Different sources need different clocks.
FreshContext uses source-specific decay rather than a universal freshness timer. The model is explicit: the assumptions can be inspected, challenged and changed.
| Source class | Typical behavior | Why one TTL fails |
|---|---|---|
| Market / price | Can move in seconds or minutes | A day-old result may be unusable even if highly relevant. |
| News / operational status | Changes over hours or days | Recency matters more than it does for stable reference material. |
| Software documentation | Usually stable until a release or API change | A very short TTL wastes good context and repeated retrieval cost. |
| Definitions / stable reference | Can remain useful for long periods | Age alone is a poor proxy for usefulness. |
What changes in the pipeline
Keep the retriever. Add judgment before generation.
FreshContext evaluates caller-provided candidate context; it does not need to own your vector database, embeddings or model.
Score freshness
Apply temporal decay relative to the source class and known timestamps.
Carry confidence
Preserve date confidence, source information and failure state in the decision.
Return a decision
Use, cite, refresh, verify, background, watch or exclude — with reasons rather than a bare score.
Leave evidence
Where ledger-backed verdicts are used, the decision can be signed and independently verified offline.
Commercial path
Start with one RAG workflow.
A bounded integration can be tested against representative queries and agreed acceptance criteria before any broader deployment.