# bert MCP server

> Provides a local hybrid retrieval layer for AI coding hosts to manage project memory that exceeds the model's context window.

- Repository: https://github.com/harshithkantamneni/bert
- Page: https://hostedmcp.org/servers/harshithkantamneni/bert
- Transport: —
- Stars: 0
- License: MIT

## Environment variables

| Name | Required | Secret | Description |
|---|---|---|---|
| `GROQ_API_KEY` | no | no |  |
| `NVIDIA_API_KEY` | no | no |  |

## FAQ

### Does bert require an LLM or API keys to function?

No, the core retrieval and memory layer operates locally and does not require an LLM or API keys. Your host model handles the reasoning.

### How does bert handle projects that are larger than the model's context window?

It uses hybrid retrieval (dense vector, BM25, and cross-encoder reranking) to provide relevant information, maintaining accuracy even when full-context stuffing is infeasible.

### What happens to model accuracy when the corpus exceeds the context window?

Without retrieval, full-context accuracy collapses (e.g., from 0.90 to 0.08), but bert's hybrid-RAG maintains high accuracy (0.90) in these scenarios.
