Kepler vs Hebbia

For investment firms that need complete lists across the whole firm, figures that come out identical on every run, and work that starts before anyone opens a project, Kepler is the stronger choice. We read every document once, in full, resolve it into your firm’s ontology, compute every number in code, and run agents when new documents arrive. Hebbia’s Matrix puts a model to work on the documents you choose, one question at a time, and lays the answers out in a grid.

We build Kepler. Facts about Hebbia come from Hebbia’s own site and public reporting as of November 2026, linked inline.

The short version

• Kepler is enterprise intelligence for investment firms, built on a living model of your firm. Models we train read every page at ingest, and code produces each figure with its formula and citations.
• Hebbia is an AI platform for investors, bankers, lawyers and corporates. Matrix puts documents in rows and questions in columns; Matrix 2.0 turns the grid into models, memos, decks and emails; Max is a chat-based AI team member; Headless Hebbia offers an API and MCP server.
• The difference is when the reading happens and who produces the number. In Matrix, a model reads each document when a prompt runs and answers each cell. We read everything up front, and code computes the figure.

Kepler vs Hebbia side by side

	Kepler	Hebbia
Built around	Your firm’s ontology: one record per fund, vehicle, manager, company and deal	Matrix: documents as rows, questions as columns
When documents are read	Every page, once, at ingest, before anyone asks	When a prompt runs across the grid
Where each number comes from	Pulled from the source and computed in code	A model’s answer in each cell, with citations
What a citation opens	The exact page, line or cell, checked before you see it	The source sentence, per Hebbia; Excel cell notes
Complete lists	Every matching document across the whole firm	Every match among the documents in the project
Your firm’s definitions	Written into the ontology, applied in every answer	Skills and workflows configured during setup
Same question twice	The same number, from the same sources	A new model run each time
As filed vs as restated	Either, on any date	Ask in a trial
Memory across deals	The ontology carries every deal, fund and manager forward	Projects, Project Intelligence and indexed deal history
Work that starts on its own	Agents on schedules and events, acting as their owner	Conditional columns; Max agents scheduled to run once
Public data	26M+ SEC filings with XBRL, 40M+ IR documents and transcripts, market data, news	Filings from six regions and transcripts; FactSet, S&P, PitchBook, Preqin, some via your subscription
Where it works	App, Excel, PowerPoint, Word, Claude, ChatGPT, Slack, email	App, Excel, Word; Claude, ChatGPT and Cursor through MCP
Build on it	API, MCP server, Python SDK, webhooks	API and MCP server
Permissions and audit	One access check on every read by a person or an agent; every query logged; SOC 2 Type II; our cloud or yours	Siloed customer data; dedicated tenant option

How Hebbia works

Hebbia was founded in 2020 and describes Matrix as a way to run a set of prompts across as many documents as diligence requires (Hebbia). You pick the documents, add a column per question, and an agent answers each cell with a citation. Hebbia’s engineering post describes chunking each document, finding the parts that need the model’s full attention, and assembling the single-document answers into the grid (Hebbia).

In 2026 Hebbia added a lot around the grid. Matrix 2.0 (August 26) carries the work through to models, memos, decks and emails, with a sign-off checkpoint before each next step, and Hebbia’s former investors and bankers index a firm’s deal history and processes during setup. Max (July) is a chat-based AI team member that works over email and has kept memory across conversations since September. The Excel plug-in puts citations on individual cells, a Word plug-in arrived in September, columns can run only when a prior column meets a condition, and Headless Hebbia (October 1) exposes Hebbia through an API and MCP server.

Data includes filings from the US, Europe, the UK, Canada, Australia and New Zealand, transcripts and investor presentations, and partners such as FactSet, S&P, PitchBook, Preqin, Daloopa, ICE and Third Bridge, some of which connect with your own subscription (Hebbia integrations). Hebbia keeps customer data siloed and offers a dedicated tenant (Hebbia security), and Sacra reports seats at $10,000 (Professional) and $3,000 to $3,500 (Lite). For how it compares with a general assistant, see Hebbia vs ChatGPT.

What users say: a consultant on Wall Street Oasis called Hebbia “genuinely interesting for large doc set extraction,” then added that “where it gets tricky is the unstructured nature of most VDRs,” because “half the critical stuff is buried in footnotes or appendices.”

How Kepler works

Frontier models keep improving and still have two gaps an investment firm feels every day. They write numbers rather than pull them, so a figure can be hallucinated and nothing flags it. And they don’t know your firm: which funds you hold, which vehicles hold them, who the key persons are, what you mean by net IRR.

We built Kepler around a different order of operations. Our forward-deployed engineers, who built this layer before at Palantir, connect your SharePoint, OneDrive, Box, Egnyte, Outlook, Slack, CRM and data rooms read-only. Before anyone asks a question, extraction models we train read every page of every document, scanned PDFs, tables and figures included, and store each fact with its value, scale and unit next to the passage it came from. Reading at ingest with our own models costs roughly 700 times less than sending each document to a frontier model, which is why we can afford to read everything. Public data is already in place: 26M+ SEC filings, 40M+ investor relations documents and earnings call transcripts, market data and news for 14,000+ companies.

Then we build your firm’s ontology. Every fund, vehicle, manager, portfolio company, deal and person resolves to one record. Your definitions are written down once. Every link is a sourced claim pointing to the sentence it came from, and the ontology knows time, so it answers as filed, as restated or as of any date. It updates itself as new documents land and gets richer every week the firm uses it. You can see how it’s structured on our ontology page.

Numbers come from code. The engine pulls each input and computes the result, every calculated figure opens as a small sheet with its formula and cited inputs, citations are checked against the source while the answer is written, and when two sources disagree you see both. Agents run on schedules and events, overnight if you like, acting exactly as the person who set them up and never taking instructions from what they read. The firm’s playbooks, such as the quarterly manager review or the IC memo in house style, are saved once and applied every time, and frontier models handle the question and the write-up while our own models do the reading. Analysts reach all of it from Excel, PowerPoint, Word, Claude, ChatGPT, Slack or email, and engineers through our API and MCP server. We hold SOC 2 Type II and deploy in our cloud or yours.

“Our analysts used to spend more time verifying AI outputs than generating them. With Kepler, every number is already traced to source.” (Senior Managing Director, Top-10 Global Asset Manager)

Not the right fit if: you’re a consultancy or law firm reviewing client documents project by project, with no funds, holdings or coverage list of your own to model.

The same question, two ways

Take a question a fund-of-funds team asks every quarter: “Every fund agreement in our portfolio with a key-person clause, and our look-through exposure to one company across every vehicle.”

In Hebbia, a team builds a Matrix with the LPAs and side letters as rows, pulled from SharePoint or the data room, and adds a column asking for key-person provisions. Each cell is a model’s answer for that agreement, with a citation. The list covers the agreements that made it into the grid. For exposure, a second grid over the latest quarterly reports pulls each fund’s position in the company, and a further step, in Matrix 2.0’s model output or in Excel, turns the column into a total. In a trial, check that every agreement is in the grid, how scanned side letters read, and whether the total matches on a second run.

In Kepler, the agreements were read when they arrived, so the key-person clauses already exist as facts, each tied to its fund and its manager in the ontology. The list comes back complete, with the named key persons and a citation to each clause. For exposure, the ontology already links the company to its position in each fund’s latest report and links each fund to the vehicles that hold your commitment. Code computes one number, broken down by vehicle, with every input cited to its capital account statement or quarterly report (what look-through exposure is). When a new quarterly report lands, an agent reruns the calculation and tells you what changed.

When to choose Kepler

• You need firm-wide answers, across every fund, deal and manager, beyond the documents loaded into one project.
• Lists have to be complete: every agreement with a key-person clause, every contract with a change-of-control trigger.
• Figures go to an investment committee or to LPs and must reproduce run to run, with the formula and sources shown.
• Your firm’s definitions (net IRR, adjusted EBITDA, exposure) should apply the same way everywhere.
• You’re an allocator, LP or fund-of-funds and need look-through exposure computed across vehicles, cited to each statement. Our guide to how LPs use AI on fund documents walks through that work.
• You want agents that act on new documents (a capital call notice, a quarterly letter, a new 10-Q) without someone opening a project.
• Restated figures matter to you, and you need to know what was reported on a given date.
• Your team wants the same cited answers inside Claude, ChatGPT, Excel and Slack.
• Your engineers want to build on your firm’s ontology and a deterministic engine through an API and SDK.
• You want engineers who built this layer before at Palantir on site in week one, connecting your sources and building your ontology.
• Compliance needs one permission check on every read, every query logged, and nothing an agent reads taken as an instruction (security).

When Hebbia may fit

• A deal team running a one-off review of a defined data room, where seeing each document’s answer in a grid is the point.
• A team that already licenses FactSet or PitchBook and wants those feeds next to a document grid.

Next steps

For the wider field, see Hebbia alternatives, AlphaSense vs Hebbia, Kepler vs V7 Go for document extraction, and Kepler vs Rogo. If the ontology idea is new, start with what an ontology is for an investment firm.

Ask Kepler about your own portfolio at app.kepler.ai, or book a demo and we’ll run the key-person and look-through question above on your documents.

Frequently asked questions

What is the difference between Kepler and Hebbia?

Kepler reads every page once at ingest, resolves everything into one model of your firm, and has code produce each figure, cited to its line, so lists are complete across the whole firm and figures come out the same on every run. Hebbia's Matrix takes the opposite order: it runs your questions across the documents in a grid, with a model answering each cell.

Is Kepler or Hebbia better for fund-of-funds and LPs?

Kepler. Allocator work depends on connecting documents across managers, funds and vehicles: look-through exposure computed across every vehicle with each input cited to its statement, every agreement with a key-person clause as a complete list, capital calls reconciled. Our engineers build that model of your portfolio as your firm's ontology. Hebbia suits reviewing a defined set of documents in a grid.

Does Kepler have something like Hebbia's Matrix?

You can lay out any set of companies, funds or documents against any set of questions in Kepler, and because every document was read at ingest, the grid fills from facts already extracted and cited. Numbers in it are computed in code. Agents keep the grid current when new documents arrive, and it exports to Excel with every cell traceable.

Does Kepler have an API and MCP server like Headless Hebbia?

Yes. Headless access is part of the platform: the Kepler API, MCP server, Python SDK and webhooks, plus Kepler inside Claude, ChatGPT, Excel, PowerPoint, Word, Slack and email. Your engineers can build agents and apps on Kepler's data, ontology and engine, and every number they get back still carries its source.

Is a Hebbia Matrix the same as an ontology?

No. A Matrix is a grid of answers for one set of documents and questions. An ontology is a standing model of the firm: every fund, vehicle, manager, company and deal as one record, linked to the others, with each link citing the sentence it came from and each fact dated as filed or as restated. Our engineers build one for every firm we work with.