WHAT IT LOOKS LIKE

One line in,
an answer with its sources out

There is no separate app to open. Typed into work chat or into a terminal, the same question takes the same path.

Work chat · terminal — same question, same answer

"Who fixed PROJ-1234?"

RouteAn issue key is present, so it is matched by exact token (FTS5) and that issue is pulled up directly

HopFrom the issue to the person, from the person to the code they raised — by relationship (KG)

AnswerWho fixed it, what they raised, and the sources, in one reply

Issue
Code review
Worklog

This is company material, so the content of the answer is masked — the question, the path it takes and the way sources are attached are unchanged.

WHY IT EXISTS

Eight places, and you are the merge step

Buying one more tool did not fix it. What was removed is not a tool but the manual step a person did in between.

BEFORE — EVERY SINGLE LOOKUP

Open eight windows
Type the same query eight times
Reconcile the scattered results by eye
Give up on questions that cross tools

NOW — MYCELIUM

Ask once, in one place
The retrieval method is chosen to fit the question
One answer comes back with its sources
Questions that cross tools resolve on the graph

Each tool's own search sees only itself — "who filed the PR that fixed this issue" has no answer there. Commercial AI search holds none of the company's data and stays generic.

WHERE THE NAME COMES FROM

The mushroom is only the part above ground

The organism itself is mycelium — a web of white threads under the soil that links separate trees root to root and moves water and nutrients between them. Here the trees are mail, issues and documents, and the threads are the relationships that connect them.

HOW IT RUNS

Collect, connect, answer

1 Collect All eight sources are swept automatically every night before dawn, taking only what is new. Nobody uploads anything.
2 Connect People, projects, issues and files become nodes, and the links between them become edges. Questions that cross two tools are answered on that web.
3 Answer The question is classified first, then searched with whichever of the four methods fits. They do not all run every time.

FOUR WAYS TO LOOK

Different questions need different reach

This is where a single method failed. Meaning search alone is weak on relationships; narrowing by the graph alone can never surface what the graph does not hold.

By meaningVECTOR

Different words still match if the sense is close — "that competitor write-up from a while back" turns up a document whose title contains none of those words.
By relationshipKG

Traversal from node to node — "who filed the PR that fixed this issue" is the kind of cross-tool question that resolves here.
By countingSQL

"The last five", "December only", "how many" — narrowing by number and date. Counting is arithmetic, not a model.
By exact tokenFTS5

An issue key such as PROJ-1234 is matched character for character and pulled up directly.

WHAT IT CAN HOLD

Eight sources, one index

Connectors are chosen to fit the site. Eight source types are connected today, and more can be added.

Work mailOutlook, Gmail and other company mailboxes
Issue trackerJira and the like — issue bodies, comments and logged hours
Document toolNotion and the like — one page is one document
Code reviewBitbucket, GitHub and the like — what was raised and what reviewers wrote
Source codeClass and method names with their locations — Java and C++ are parsed structurally
Local filesPDF, HWP and similar — whole folders are swept
Product manualsDocumentation published to the web — one page is one document
Company policyRules and notices held in the groupware system

Scale differs per site — 4–12s measured at around 50,000 documents

Where the material sits and who can see it

What it can seeIndexed under your own account — only material you already have rights to
Your own materialHeld in your index alone, never mixed into the department one — linked in only when you search
Where it runsCloud, on-premises or a local server — the index and material stay where you install it
What leaves the networkEmbeddings and relation extraction only — everything else runs inside the installation

WHAT PEOPLE ASK

Plain sentences in, one answer out

You do not have to phrase it carefully. Ask in the sentence you would use anyway.

"What is the travel expense policy?"

Finds the clause in the company policy held in the groupware system
"Who would know about this?"

Proposes people from worklog, pull request and issue history
"Book a room on floor 11 at two."

Picks a free room and completes the reservation

LIMITS

What it deliberately cannot do

The tool never invents an answer outside its range. When nothing is found it answers not confirmed within the search scope, rather than filling the gap.

Source code bodiesOnly class and method names and their locations are indexed — implementation is read from the file
External materialISO standards and outside white papers are not held
Scanned PDFsAn image-only document has zero text, so the file has to be opened directly

IN TECHNICAL TERMS

Hybrid retrieval over a knowledge graph

The same thing said in industry terms.

Graph: People, teams, projects, issues, documents, tools and code symbols are nodes — owns, created, depends on, references and decided are edges
Semantic layer: Every source mapped onto one type system — one vocabulary in place of many
Retrieval: Four methods with query routing — vector · full-text (FTS5) · SQL aggregation · graph traversal
Answers: RAG with citations — sources come back alongside the answer
Indexing: Incremental and cursor-based — nightly, new records only
Storage: SQLite and a vector store — installs on cloud, on-premises or a local server

External models are used only for embeddings and relation extraction — everything else runs inside the installation

MEASURED

Every figure comes from daily use

Each figure carries what was measured and when.

98.1%measured on a 417-question regression set grown from real questions2026-05-20
4–12sone-line question at ~50,000 documents; slowest seen 45s2026-09-03
$0.5/dayembeddings and relation extraction on an outside model2026-09-03

The index has grown since that scored run; the old figure stays until it is measured again. Scale differs from site to site, so per-installation counts are never used as figures.

Eight windows become one line

You ask from the work chat and the terminal you already use. There is no new screen to open.

See Dispatch Open the Dispatch demo Dispatch in detail 한국어

One line, askedacross scattered material

One line in,an answer with its sources out