FELLOW
Platform

Stored context, made visible.

Behind the scenes, FELLOW hands bounded, cited context to a model, not a bag of similar chunks. This page is the other half: what that same context looks like when a human can read it, and correct it, directly.

01. Beyond RAG-only

A map for truth. A search net for scale.

Vector retrieval is excellent at “find something like this.” It is a bad design choice for “what is our canonical definition,” “who reports to whom,” or “what stage is this buyer in.” FELLOW is built as a hybrid: governed structure for high-stakes claims, retrieval for exploration.

Dimension RAG-only pipeline FELLOW
Core unit Fragmented text chunks in a vector index Atomic facts with evidence, linked into shapes
How answers are found Probabilistic nearest-neighbor similarity Structured + graph paths first; semantic when you explore
Structure Chunking often flattens tables, stages, and reporting lines Org charts, process maps, and personas assembled from links
Staleness An outdated but similar chunk can rank highest Supersession and provenance; current state is explicit
Human interface Query a black-box index; hard to correct in place Readable views; edits flow through the same review path
Honesty about absence Missing context is usually silent (or hallucinated) Known gaps are first-class and actionable
Best fit Wide, fuzzy search over archives Canonical account knowledge agents and teams act on

Hybrid in practice

Asking a probabilistic system for your company’s high-stakes commercial truth is the wrong job for that tool. Keep search for scale; put reviewed structure under decisions. More in the blog →

02. Visualizations

The same context, three shapes.

FELLOW doesn't have one generic viewer. It renders each record type into the form a person actually expects (an org chart, a process map, a profile), never a raw JSON blob.

org_chartsrc: fct_04J1
Gavin BelsonVP Sales
Richard HendricksSales Manager
Dinesh ChugtaiSales Manager

Reporting lines assembled from graph-linked facts ("Richard reports to Gavin," "Dinesh reports to Gavin"), not a synced org-chart import.

Correct a reporting line here; it's promoted back through the same review path as any other edit.
buyer_process_mapsrc: exp_01J8Q
Aware
Evaluating
Deciding
Expansion

Currently: pricing objection raised on last call, procurement now looped in. Stage advanced automatically when that objection was logged.

A buyer's stage isn't a picklist a rep sets by hand; it moves when the underlying evidence changes.

personasrc: exp_01J77
EB

Erlich B.

Economic buyer · Champion

  • Prefers monthly retainer over fixed-scope project[1]
  • Skeptical of "vapor," wants proof before commitment[2]
  • Reports to a historically cost-sensitive buyer[3]

Every attribute is a cited assertion, not a free-text CRM note: click through to the source conversation.

Add a fact here; it's queryable immediately, flagged for review only if it conflicts with something already known.
03. Underneath

Every shape is built from cited facts.

The visualizations aren't the data; they're a view of it. Underneath each one is a set of individual, traceable assertions.

  fct_04J1: cited factauthority: contractual

"Richard Hendricks reports to Gavin Belson (VP Sales)."

Traces to

  • kickoff-call · 2026-04-02 · 00:12:44[1]
  • org-announcement.pdf: p.2[2]

State

  • current · supersedes fct_03X8 (Q4 structure)
Cited, traceable facts

Traceable back to the moment it was said.

Every node in a chart, every stage on a map, every line on a profile is one assertion with a source attached. Follow it back to the exact clause, timestamp, or workbook cell, and see what it superseded. In the spreadsheet pilot, each record kept its cell origin and version history, so reviewers could inspect the numbers behind an answer.

  • Source-linked at the level of the individual fact
  • Supersession tracked, so nothing stale is shown as current
  • Corrections flow back through the same review path
  acct_4471: known gapsdiscovery

What we don't yet know

  • Signing authority above $50k: unconfirmedgap
  • Incumbent vendor & contract end dategap
  • Security review requirement: implied, unverifiedgap
Known gaps & discovery

And it shows you what's still missing.

A profile with holes shouldn't look complete. FELLOW tracks the questions it can't yet answer for a record and surfaces them, so the next call, or the next agent run, knows exactly what to go find.

  • Missing facts flagged rather than silently guessed
  • Gaps become a concrete discovery checklist
  • Agents can request precisely what's absent
Pilot proof

From messy workbooks to a cited fact graph.

A figure needs a period, a unit, a current version and a source. In a pilot for a mid-sized services company, FELLOW connected workbook records to people, plans and the cells behind the numbers.

01

Capture and type

FELLOW inventoried Excel workbooks and hidden tabs, retained formulas and reported values, and recognised layouts from headings. Named columns separated currency, percentages and counts. Unreadable sources stayed visible.

02

Resolve and connect

FELLOW matched names to employee records and linked plans to periods, then tracked repeated rows across versions and separated totals from detail. Ambiguous matches and inferred plans stayed labelled.

03

Audit and test

FELLOW checked every sheet's outcome and measured captured numbers against loaded columns. Both years passed automated testing, with at least 99.95% of captured numbers in named columns.

04

Retrieve the evidence

Current-version views support totals. Linked records compare statements, requests and approvals. Each figure traces to the workbook, tab and cell it came from: 239 of 240 sampled values matched their source cells.

The pilot made the relationships queryable in one database: figure, period, person, plan and source. Open questions remained part of the record.

Try the trace on anonymised pilot data →

Bring your workbooks to a design-partner discussion →
04. Roadmap

What's shipping next.

Built in the open, in dependency order: each phase unlocks real usage, not a demo.

Phase 0: Retrieval hardening In progress
01

Multi-channel access

Query and update FELLOW from where work already happens, starting with Slack, using per-person keys so every request is tied to a real identity, not a shared token.

02

Auto-capture from conversation

A session with an agent or a colleague turns directly into knowledge: extracted, scored, and immediately usable, with review reserved for the cases that actually need a second look.

03

Visual interface

This page is the preview: the same read surface shown above, plus the ability to correct, add, or promote a record without leaving the view it was rendered in.

04

Broader structured ingestion

Workbook ingestion is available and proven in the pilot. Next: long, structured source documents (filings, contracts, policy manuals) decomposed into linked, queryable records, with the full source always one reference away.