A graph of your documents, not a pile of chunks. Accurate retrieval for your AI.
Nothing to chunk or tune: TGM organizes every file on its own. Each passage comes with its file and version as proof. No LLM calls, so no LLM bills.
Illustrative example
- A pile of chunks: pieces of three files in a messy heap, one cut mid-sentence, none showing a file or version.
- TGM organizes them into a graph: one cluster per file, short files whole, the contract as passages cut at sentence ends, its versions stacked behind it.
- Your AI asks: “How much notice do we need to cancel?” The right passage lights up.
- The answer: that passage lifts out whole, with services-agreement.md · version 2, and arrives at your AI.
“Where did that answer come from?”
Without TGMA chunk of text, maybe cut mid-sentence, with no version.
With TGMThe whole passage, with its file and version.
How it works
- Add your files. You get a receipt for each one.
- TGM builds the graph. No chunk sizes, no settings.
- Your AI asks. TGM finds the best passages, then looks once more.
- Your AI gets proof. Whole passages, each with its file and version.
Illustrative example
- Add your files. You get a receipt for each one.
- TGM builds the graph. No chunk sizes, no settings.
- Your AI asks. TGM finds the best passages, then looks once more.
- Your AI gets proof. Whole passages, each with its file and version.
“How long does setup take?”
Without TGMChoose a chunk size and overlap, test, rebuild.
With TGMAdd your files. Nothing to tune.
Your files change. TGM keeps every version.
Ask about any past day and get the same passages your AI got then.
Illustrative example
- Your contract changed in June. TGM kept both versions.
- Ask today: sixty days, from version 2.
- Ask about March 3: thirty days, from version 1.
- Ask again any time: same question, same day, same passages.
“What did our AI see in March?”
Without TGMThe June change wiped out the March wording.
With TGM“thirty days’ written notice”, version 1, word for word.
When one fact leads to another, TGM looks twice.
The first search finds one fact. The second look uses that fact to find the next one.
On HotpotQA, a public test of 7,405 two-step questions, every passage TGM needed was in its top 20 results 98.4% of the time. Plain search on the same documents: 94.8%.
Illustrative example
- A question that needs two facts.
- The first search finds who the biggest customer is, but no notice period. Plain search stops here.
- TGM takes a second look: the question plus that passage, searched once more.
- It finds Northwind's notice clause. Your AI gets both and can answer: sixty days.
“What's the notice period with our biggest customer?”
Without TGMPlain search finds who your biggest customer is, but not their contract's notice clause. Your AI can't answer.
With TGMTGM finds “Our biggest customer is Northwind Health”, looks again, and brings Northwind's clause: sixty days.
Proof, not promises.
0 settings to tune: add your files, and TGM does the organizing.
100% word for word: every passage is your file's own text, never rewritten.
98.4% of two-step questions: TGM found every passage needed. Plain search: 94.8%. (HotpotQA, a public test)
Same answer, every time: ask again later about the same day and get the same passages, word for word.
Every public result, including the side-by-side with LangChain's standard setup, is on the proof page.
What you get today
- Works with text and Markdown files
- Every version of every file, kept
- The exact words, with the file and version for each
- Hosted by us, so there's nothing for you to run
- Use your own embedding model if you want
- A Python library to connect your AI
Questions
How do I get an account?
Book a call. We set up your account after it and email you a link to log in.
Does TGM write the answers?
No. It hands your AI the exact words from your documents. Your AI writes the answer.
Which files work?
Text and Markdown today. TGM doesn't read PDFs or other formats yet. We convert them with you when you start.
Where does it run?
We host it, so there's nothing for you to run. You can use your own embedding model.
SOC 2, permissions, or your own cloud?
Not yet. Tell us what your security team needs, and we'll go through it on the call.