Most retrieval-augmented generation setups begin with chunking: documents are sliced into fixed-size pieces, chopped somewhere mid-sentence, and stored as anonymous fragments. SearchCandy Labs argues this is where accuracy dies. Its product, TGM, takes a different route — it treats each file as a unit, builds a graph of your documents, and keeps short files whole while cutting longer ones at sentence boundaries. Vertically, it stacks every version of a file so the system can answer questions about a specific point in time.
The result is meant to be proof rather than probability. When your AI asks for something, TGM returns complete passages accompanied by their filename and version, and the vendor claims you can re-ask about a past date and receive exactly the same passages your model saw then. Because the system does not call an LLM to do the retrieval work, there are no per-query LLM bills. TGM also runs a two-step lookup for chained questions: the first search finds one fact, then the query is re-run with that finding so a second fact can be located. On HotpotQA, a public 7,405-question benchmark of two-step questions, the vendor states every required passage appeared in TGM's top 20 results 98.4% of the time.
Deployment is hosted — there is nothing for you to run — and a Python library connects your AI stack to it, with the option to bring your own embedding model. Access is arranged through a sales conversation: the site's primary call to action is to book a call, after which an account is provisioned and a login link is emailed. There is no public self-serve pricing table, and the company publishes a proof page documenting its retrieval results.
