AI & SYSTEMS

My research tools returned exact matches alongside related passages

Keyword and meaning and graph searches produced one ranked list with the retrieval method shown.

One question is asked by exact words, by meaning and by connections at once, and comes back as one ranked list that shows what found each result.

An AI agent's question splits through three lenses: exact words, meaning and connections. Each lights matching records on the floor, and connections draw links between records. The matches fly into one ranked list, each row wearing a badge for every lens that found it. — One question is asked by exact words, by meaning and by connections at once, and comes back as one ranked list that shows what found each result.

search modes, one ranked list

platforms on my own systems

Public

trail at WikiDesignCo

Private

internals, by policy

Case Studies

Context

So I build it for myself, with no client, no invoice and no deadline anyone else set.

Problem

Every agent system inherits the quality of what it can retrieve. The off-the-shelf stacks I've used answer one kind of question at a time, and a real customer's world needs connections, exact words, and meaning answered in one query. So I built a system that does that.

Approach

The Metagraph runs full-text, vector, and graph retrieval as one platform, with agents driving it. The Content Compiler treats content as code: written material as structured data that can be measured, revised, and recompiled. Both run on my own systems, and WikiDesignCo is where their public work shows up. The internals stay private.

Stack

Multimodal retrieval (full-text, vector, graph), run by agents
Content as compiled, measurable data
Public platform: WikiDesignCo
Published trail: the WikiDesignCo research articles

Result

retrieval methods in one ranked list

The platforms exist, run, and back the public writing at WikiDesignCo. Anything published there is published long after the value came out of it.

Impact

When one of my agents searches, it gets the exact string, the paraphrase and the linked record back in one ranked list, with a badge on each result for the search that found it.

Lessons

I keep what I learned here private, the way a quant firm keeps its models: the world can know what the system does, and the internals stay home.

Who this is for

For anyone checking whether I build agent infrastructure or only talk about it. The platform is live at WikiDesignCo.

Get a quote

See it live

See the live platform(opens in a new tab)

On this page