What should a knowledge base give an AI system besides text?
It should give the system three things: passages it can retrieve, tools it can call, and a way to check its answers. A physics textbook project, StringQFT, is built around that pairing. Its abstract describes the design but reports no performance results, so it is a useful pattern rather than evidence of better answers.
At a glance
- Finding: The project packages one textbook as retrieval records, calculation tools and benchmark questions.
- Why it matters: The AI is handed specific, traceable context for a defined task, not a pile of documents.
- Limit: Only the abstract was reviewed. It is one physics resource and shows no outcome data.
What the project contains
According to the paper's abstract (posted to arXiv in November 2026), StringQFT is an 18-volume textbook collection with 190 chapters. It is presented both as a textbook and as infrastructure for AI-assisted theoretical physics.
The package includes several parts for AI use:
- 1,371 section records linked to textbook sources and PDF locations, plus a local lexical retriever (keyword-style matching) for retrieval-augmented generation.
- 19 calculation tools with specified inputs and finite domains.
- A benchmark of 1,073 open-book items, 499 of which have automatic finite-answer checking.
- 21 private, source-linked examples for preparing training data.
Why source links matter for context
Context engineering is about giving a model the right evidence for the task. A section record that points back to a textbook location lets a person check where a retrieved passage came from.
An illustrative example: a model asked about a brane construction retrieves a section record. A reviewer can open the linked PDF location and compare the answer with the source. Without the link, the reviewer would have to search the whole book.
The abstract does not say how often retrieval finds the right section. That would need its own test.
Why tools with defined limits help
The abstract says each of the 19 tools has specified inputs and finite domains. In plain terms, the tool states what it accepts and where it works. That is a context decision: the agent gets a bounded calculation instead of being left to improvise one.
The abstract also says the tools connect to textbook derivations. That suggests the answer, its source and the calculation can be read together. The abstract does not report whether agents use the tools correctly.
What remains uncertain
The abstract illustrates the uses with specific questions, passages and calculations. It does not report accuracy, retrieval quality or a comparison with other setups. A lexical retriever may suit a technical textbook with precise terminology, but this source does not show that it beats other methods.
The pattern also comes from a narrow, well-structured domain. Marketing sites, support content or product catalogs may not have answers that can be checked automatically.
What to test next
- Pick one task with a checkable answer, such as quoting a policy or computing a price. Link every retrievable chunk to its source location.
- Define each tool's allowed inputs and ranges, then log cases where the agent goes outside them.
- Compare answers with and without the retrieved records. Measure correct answers, correct source choice and unsupported claims.