FOR DEVELOPERS
Predicate pushdown,
for AI retrieval.
Aspected is a database built around one patented invention: a multi-aspect query index. Every metadata signal — type, year, region, classification — becomes an aspect of the query itself, graded alongside meaning in one scoring pass. Nothing is filtered away before scoring; nothing needs repairing after. And every score is decomposable, per signal.
● US Patent · PCT filed ● MCP-native ● Runs alongside Pinecone, pgvector, Weaviate & co. ● One Docker container ● Your internal data — never the public web
01 · WHY ADD IT
Three things change in your stack
Inspect & control, at retrieval
A dense vector gives you one blended similarity number — observable, not inspectable. Aspected scores every aspect separately in the same equation, so every ranking decision can be opened, read, and changed. Weights are dials, not code: tune, re-run, verify — in place, no re-chunking, no pipeline rebuild. Save the winning dials as a retrieval policy: versioned, auditable, steerable per query. Trust isn’t claimed; it’s checkable.
Fewer tokens, straighter path
The patched pipeline pays three times per question: an oversized shortlist to survive the filter, a reranker to repair the order, re-query loops to recover what the filter discarded. One pass over all the signals means the loop never starts — and there is no cutoff: nothing is deleted before scoring, so ROT content demotes itself instead of polluting the shortlist. The right context in, verified, the first time. The agentic layer keeps its own checks; it simply has far less to do. Right the first time, for a fraction of the tokens.
No knowledge graph to build
Instead of metadata as a filter — or an entity graph you must extract and keep in sync — Aspected uses metadata as a signal. Most enterprise queries are single-hop-plus-constraint: the current, authoritative, permitted document about X. That is what most teams reach for a graph for — structured signals shaping the ranking, with an explanation. Aspected delivers it from metadata your files already carry: no LLM extraction pipeline, no entity-resolution upkeep, nothing to keep in sync. The honest boundary: for arbitrary multi-hop traversal or corpus-wide summarization, a graph still earns its keep — keep one for that, and route those queries to it.
02 · QUICKSTART
Add a database. Don’t rebuild a stack.
Nothing is migrated, nothing is decommissioned. You add one container, compare on your own content, and route only what wins.
1 · Run the container
One Docker container beside your current stack. It exposes an MCP server any agent can reach — a Bedrock Agent, Copilot in a Microsoft tenant, or your own orchestration.
2 · Turn metadata into aspects
Point it at your internal content — Aspected answers from the sources you connect, never from the public web. Every metadata field your files already carry — type, year, region, classification — becomes an aspect, graded alongside meaning. No taxonomy project, no cleanup phase. Every vendor’s deal is clean or pay — with the signals in the ranking, cleaning becomes a choice.
3 · Compare, then route
Run your ten worst queries against your current retrieval and Aspected, side by side — about thirty minutes. Over MCP, your agent then routes each question to the method that answers it best. The results do the arguing.
03 · INTEGRATIONS
One MCP server. Any agent.
Amazon Bedrock
A Bedrock Agent reaches Aspected over MCP: Bedrock Agent → MCP → Aspected — including inside your own VPC.
Microsoft Copilot
In a Microsoft tenant, Aspected is deployed as an MCP server and the Copilot agent routes questions to it — your users stay in the Copilot they already have.
Your own orchestration
Straight MCP from your own stack, or a sovereign interface like Nextcloud’s assistant. Runs alongside the vector store you already have. Coming in September: LangChain and LlamaIndex retriever integrations.
04 · TRY IT
Debug your ten worst queries.
The 30-day free trial is the full Aspected Database in your own account — your own internal content, never the public web, full dials, up to 5 TB. Two ways to fill it: bring raw documents and the prep pipeline takes care of the fundaments (parsing, chunking, and turning the metadata your files already carry into aspects) — or push your own records through the ingestion API, indexed as they are (1 TB ≡ 40 million records). And from September 7, start it straight from AWS Marketplace — listing: Aspected Database — no sales call required.