cleanup uc0.1#8
Open
Rahien wants to merge 12 commits into
Open
Conversation
…down to a minimum and keeping the info in the ai components. added placeholders for PIR, added a word about why we don't do codlist management tools
…ure-work uc2 with sparql future work
brechtvdv
reviewed
Jul 16, 2026
Collaborator
There was a problem hiding this comment.
The file mentions "4. Store the result. The impact direction is stored as an ext:has_impact predicate on the annotation in the AI graph of the triplestore.". However, ext:has_impact does not exists as the impact stored as a second oa:hasBody. Can you update?
brechtvdv
approved these changes
Jul 16, 2026
|
|
||
| - Users with malicious intent could convince the LLM to run destructive queries, inserting or deleting triples into the SPARQL endpoint. This is easily countered by leveraging the [capabilities of mu-authorization](https://github.com/mu-semtech/sparql-parser#define-access-rights-for-specific-services) Simply giving the LLM a scope that only allows read access to the public information in the triplestore is enough. | ||
| - Other ill-meaning users could decide to have the LLM run a lot of very heavy queries, resulting in a Denial of Service attack on the system. This can be mitigated by 1) providing the LLM with patterns of such malicious queries and telling it to refuse executing them, 2) extending mu-authorization so it disallows queries matching such patterns, 3) putting limits on the execution time of queries (built in for virtuoso) or queries sent from a certain service (requires an extension of mu-authorization). | ||
| - Because the LLM now operates as an agent that can take actions, the risk of prompt injection where malicious content in a retrieved document manipulates the model's behaviour. This becomes more consequential than in the current stateless pipeline, however the plan-execute architecture partially mitigates this by separating planning from execution and by running the monitor as deterministic code rather than as an LLM call, but careful prompt engineering and input sanitisation remain important concerns. |
Collaborator
There was a problem hiding this comment.
Rewrite: the risk of prompt injection...
Suggestion: Since the LLM now operates as an agent that can take actions, it becomes vulnerable to prompt injection—where malicious content embedded in a retrieved document manipulates the model's behavior.
|
|
||
| A critical enabler for this architecture is providing the LLM with sufficient knowledge of the data model. Without schema context, the LLM would waste significant effort discovering endpoint structure on every request –which classes exist, which predicates connect them, which prefixes to use or worse, produce syntactically valid but semantically incorrect SPARQL. | ||
|
|
||
| The proposed solution is to supply SHACL shape files and example queries as retrieval context before planning begins. SHACL shapes express constraints, properties, prefixes, and class structure in a machine-readable format that the LLM can interpret. Example queries package ready-made patterns, comments, keywords, and endpoint information. Together, they give the planner enough orientation to produce actionable steps and give the executor enough vocabulary to construct correct SPARQL. |
Collaborator
There was a problem hiding this comment.
Move this to possible future work?
Now using void?
Contributor
Author
There was a problem hiding this comment.
this is in future work
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
commits are scoped, have a look at the commit messages for description and/or go commit by commit while reviewing