ALM: User Guide

For the AI features built into the application to work, two components are required:

• a vector database, containing so-called embeddings — vectorized representations of the texts of project artifacts. Because project artifacts change rarely, while a search for similar artifacts in order to perform AI features is required often, it is efficient to cache the vector representations rather than request vectors from language models every time. That is what the vector database does.
• a language model — on the one hand it is the source of vectorized representations, and on the other it makes it possible to implement many AI features using text processing and transformation, as well as creation of texts based on the texts of project artifacts.

When the application is deployed, an open and free language model is used by default. You can use arbitrary language models supported by the local services Ollama, vLLM, and ollama.cpp; however, keep in mind that good language models require significant and specific computing resources (for example, a GPU). An alternative is to use external language models such as GigaChat and YandexGPT. The quality of the language model determines the quality of AI features in the application.

To search for project artifacts that are similar in meaning, for the purpose of subsequent analysis or other computations, the RAG (Retrieval-Augmented Generation) mechanism is used.

In the background, ALM requests embeddings from the LLM for the texts of previously created or changed project artifacts; the result is saved in a special store required for a fast search of texts that are close in meaning.

Thus, various requests to the LLM use the texts of your project artifacts, making it possible to perform useful AI features during design, analysis of links, and so on.

Using MCP and the REST API, you can implement your own RAG scenarios; for details: AI Features

For example, a new Issue can be linked (traced) to a previously implemented Requirement and thus potentially change it. Or a test scenario (test case) can cover (be traceable to) a functional Requirement.

Traces are created automatically while working on project artifacts, for example during the lifecycle of an Enhancement (or a user story). Or traces are set manually.

In the early stages of a project, creating traces manually, for example between Requirements of different levels (types), requires a large amount of manual labor. On the other hand, project documentation changes continuously, which also leads to the need to add or change traces. To save time and reduce the number of errors in this process, ALM implements an automatic tracing mechanism (auto-traces).

Based on the results of the RAG mechanism, ALM performs automatic selection of traces between project artifacts along the key references: Requirements of Issues, test and operational documentation, and so on.

An auto-trace is not set automatically; it requires validation by the user.

When you click the "add" link, the system displays suggestions for creating auto-traces.

In the modules for working with the main project artifacts, the useful modules "Auto-trace for..." are available, where you can see the full picture of ALM's suggestions for auto-traces: