PipesHub 1.0.0 has released a platform that brings together scattered company content and makes it usable for two purposes: for internal search and as a data basis for AI agents. Connected are, among others, Slack, Jira and GitHub – precisely the systems in which the actual working knowledge lies in many development and project organisations.
According to the project, new in version 1.0.0 are code graphs as well as stricter permission checks. Both points sound technical, but they decide whether an AI assistant gives reliable answers in everyday work or represents a risk.
The basic problem: knowledge is there, just not retrievable
In medium-sized companies, knowledge rarely arises in one place. A decision is made in the chat, the reasoning is in the ticket, the implementation in the repository, the acceptance again in the chat. Anyone who later asks why an interface was built the way it was built has to search three systems – and in doubt does not know the right search terms.
This is exactly where platforms like PipesHub come in. They index the source systems and place a common search layer over them. For AI agents – that is, programs that process tasks semi-autonomously and look up information for this purpose – this layer is the prerequisite for being able to work with internal context at all instead of only with the general knowledge of a language model.
Code graphs: structure instead of text snippets
Common search methods for AI assistants break documents into sections and find similar passages of text. For source code, this only works to a limited extent. A function is rarely understandable on its own; what is relevant are the call relationships, the dependencies and the place where a value is originally set.
A code graph maps such relationships as a network of nodes and edges – files, functions and their connections to one another. For questions such as "Which places are affected if we rename this field?", this is the more suitable representation than a list of similarly sounding code fragments. For project teams, this is above all interesting for legacy systems in which no one has the complete overview any more.
Permissions are not an afterthought
The second point weighs more heavily for decision-makers. As soon as content from several systems lands in a common index, it leaves the protected space of its source application. A private project channel, a ticket with a personal reference, a repository with customer code: In the source, permissions apply there that are simply no longer present in the search index.
If this index is then connected to an assistant, an inconspicuous kind of permission escalation arises. No one breaks into a system – the information simply becomes accessible via a new path. That PipesHub tightens up on permissions with version 1.0.0 is therefore not a marginal detail, but the condition for such systems being allowed to be used in regulated environments at all.
In practice, this means: The permissions of the source systems must be enforced at query time, not only at import. Anyone who is not allowed to see a ticket must also not receive an answer that was summarised from this ticket – and no reference to it in the source list.
What needs to be clarified before the pilot project
Regardless of the specific tool, sober preparatory work is worthwhile before an introduction:
- Source selection: Which systems actually contain decision-relevant knowledge – and which above all noise? Not every chat channel belongs in a knowledge index.
- Permission model: Are the permissions in the source systems maintained at all? A search index makes visible every imprecision that has grown over years.
- Data categories: Where is personal data, where is customer material with contractual conditions? These areas need a conscious decision, not a standard configuration.
- Operating model: Who operates the index, where is the data located, how are deletions from the source systems followed up?
- Traceability: Does the system provide evidence for its answers? Without a source reference, the quality cannot be checked.
Quality arises before the model
In projects, the quality of AI assistants is frequently attributed to the language model. In fact, the majority is decided beforehand: If the retrieval delivers the wrong or outdated documents, even a strong model only formulates a convincingly sounding incorrect answer. Outdated concept papers, rejected solution proposals from old tickets or code states from abandoned branches are the typical sources of error here.
That is why work on the data basis is not a preliminary project, but the actual core. This includes decisions about which content is considered binding, how up-to-dateness remains recognisable and which areas are deliberately excluded.
Classification for web and software projects
For project organisations, the focus shifts: It is not the choice of model that is the critical step, but the clean preparation and safeguarding of one's own knowledge. Anyone who uses Slack, Jira and repositories anyway already possesses the data basis – it is just not retrievable and often imprecisely maintained in terms of permissions.
A pilot with a clearly defined scope of sources, enforced rights and verifiable answers is the more sensible starting point than a broadly rolled-out assistant whose information nobody can check.
Sources
- KI-Agenten an der kurzen Leine: PipesHub durchsucht das Firmenwissen — heise developer News