MARE started from a simple question: do we really need to embed and retrieve everything? Traditional RAG works well, but as datasets grow, the default approach often becomes: chunk everything, embed everything, retrieve Top-K, and hope the right context is inside that window. That creates a growing vector footprint and fixed retrieval behavior. It can also struggle when answering a question requires navigating across multiple related pieces of data rather than finding one semantically similar chunk. I wanted to explore a different architecture. MARE — Mongo Adaptive Retrieval Engine — treats retrieval as a navigation problem. Instead of maintaining an embedding for every chunk, MARE keeps a much smaller semantic navigation index and lets an agent investigate the underlying data using MongoDB tools such as search, filtering, aggregation, and direct document retrieval. The agent can progressively explore the dataset, follow relationships, resolve entities, and decide what information it actually needs before generating an answer. The project evolved quite a bit while benchmarking it. Early experiments showed that MARE could reduce the persistent vector footprint dramatically. In one 10K-incident benchmark, 604 navigation vectors replaced roughly 60,000 RAG chunk vectors — about 1%. But the experiments also exposed the trade-offs. RAG was significantly faster and more token-efficient for straightforward semantic lookup. MARE became more interesting on questions involving cross-collection traversal, multi-hop investigation, entity resolution, aggregation, and indirect clues. In one LLM-on benchmark of 20 held-out questions: • MARE: 19/20 correct • Hybrid RAG: 18/20 correct • Persistent vectors: 604 vs 60,000 One query was especially revealing. The question mentioned neither the customer nor the root cause directly. MARE followed the surrounding evidence, identified the relevant customer, traversed deployment and migration information, and eventually discovered the underlying authentication failure. Fixed Top-K retrieval missed the bridge entirely. The biggest challenge turned out not to be search. It was retrieval policy: how much should an agent retrieve, when should it stop investigating, and how do you prevent the context from exploding? That led to experiments with retrieval receipts and bounded investigation, which reduced MARE's token usage substantially while keeping answer quality intact. MARE is still an experiment, not a claim that agentic retrieval replaces RAG. The emerging idea is more nuanced: RAG is excellent at retrieving known semantic neighborhoods. Agents may be better when the answer has to be discovered. MARE is my attempt to understand where that boundary actually lies.

MARE

Adaptive Retrieval Engine for Agentic Stack

Adaptive Retrieval Engine for Agentic Stack