Your RAG Is Not Failing. It Is Reading Too Much.

Most RAG failures are context pollution failures. Retrieval still works.
The correct document is often retrieved, but the model also receives unrelated documents, stale notes, wrong-domain records, old decisions, and tool output that never belonged in the context window. That is how systems produce confident wrong answers while dashboards still look green.
The failure mode
A user asks:
What medications am I currently taking?
The retriever finds the right medical record, but it also pulls financial documents because they contain words like:
- statement
- balance
- discharge
- liability
The answer is present; the context is contaminated. Most RAG evals still count that as a retrieval success because the correct source appears somewhere in top-k. And the user sees a wrong answer.
What we measured
On a mixed Financial + Medical dataset:
| Strategy | Recall | Cross-domain noise |
|---|---|---|
| Flat mixed index | preserved | 2.5 wrong-domain docs/query |
| Sharded + Hybrid RRF | preserved | 2.9 wrong-domain docs/query |
| Sharded + learned routing | preserved | 0.0 wrong-domain docs/query |
The surprising result: hybrid fusion made irrelevant context worse. The answer was findable and polluted.
Why this matters
Most systems optimize:
Can I retrieve the correct document?
Production systems need to optimize:
What should the model not read?
Those are different problems.
This gets worse as teams add more sources, tools, MCP servers, retrievers, memories, and skills. Capability count grows, context quality falls, and the model sees more definitions, schemas, snippets, and partially relevant evidence than it needs. That is context bloat.
The remedy
Stop injecting every capability into the model context. Treat capabilities as plugins selected by the planner, with tools in the prompt only when selected. The model should not see every retriever, database schema, MCP tool, skill, and memory store; the planner should select the minimum evidence needed.
For Context Runtime, that means:
- retrieval methods live in a registry
- data sources live behind plugins
- memory indices are selected per request
- tool schemas stay out of the prompt unless needed
- the planner chooses the execution strategy before the model is called
The result is cleaner context, lower cost, and fewer confident wrong answers.
Bigger windows do not solve context pollution. Better planning solves it.
Reference implementation: https://github.com/redevops-io/context-runtime
Benchmarks: https://redevops.io/benchmarks
Whitepaper v2: https://redevops.io/whitepaper-v2
redevops.io builds rAIlsApps (source-available (AGPL-3.0 with the Commons Clause)), self-hostable modules run by agents, on hardware you can actually afford. See it run.

