Applied AI Research · Information Retrieval · Evidence
ParancU Lab explores the limits of semantic document retrieval and investigates methods for identifying factual evidence within source material.
The work combines prototype development, controlled experiments and evaluation of both successful and unsuccessful approaches.
Explore the research directions →Semantic similarity is not answer containment. A passage can relate closely to a question without containing the facts needed to answer it. Embedding-based retrieval can find factual answers; similarity alone does not establish sufficiency.
Illustrative example, not an experimental result: a manual says, “Do not operate the pump above 60°C.” A question about whether operation at 80°C is permitted concerns the same topic, but an answer must preserve the prohibition and temperature condition.
Missing qualifications, negation, incorrect factual relationships and incomplete evidence can make relevant passages misleading. Sometimes the answer is absent from the available documents. Recognizing these boundaries is an open research problem.
Interconnected research directions
Investigates factual representations, question-oriented retrieval and alternatives to global semantic similarity, while examining source preservation, factual relationships and answer containment.
No generally validated retrieval architecture has yet been established.
Investigates evidence sufficiency, complementary passages, missing-information diagnosis, verification boundaries and targeted retrieval recovery.
Tested components remain experimental. Integration with ParancU 2.0 is under development, not a validated combined product.
The earlier agentic evolution provides an environment for investigating retrieval, AI-assisted generation, source inspection and verification workflows.
Generating an answer and verifying its support are separate challenges. Verification does not guarantee correctness.
Explore Agent Lab →Evaluation
Experiments examine whether:
Retrieval rankings and evidence-containment measures are distinct from answer correctness. High similarity or a successful isolated benchmark does not establish general factual reliability.
Findings & limitations
Research snapshot · 20 October 2026
ParancU v1.1 provided a working embedding-based retrieval baseline. Implemented newer components include source-linked sentence representations, multiple-question retrieval and experimental evidence-selection logic.
Question-based representations and alternative matching strategies have been investigated. Some small controlled comparisons produced encouraging results, but learned-matching experiments failed their success criteria and a source-binding follow-up exposed attribution failures.
Decision Lab has tested sufficiency and selection components in synthetic scenarios and selected real-document evaluations. Results include successful judgments, false sufficiency judgments and rejected sufficient evidence. Synthetic outcomes are not real-document performance estimates.
Software tests and model-generated judgments do not constitute independent validation of factual correctness. General reliability and generalization beyond the evaluated material remain unestablished.
Open questions
Cloud processing and API access are possible development directions, not established services or a committed architecture. Document confidentiality remains a design requirement; processing, access and retention arrangements require definition and validation.
Publication
The paper documents earlier retrieval architecture, evaluation and limitations. It does not validate the newer research directions.
Read the white paper (PDF) ↗ParancU Lab documents implementations, experiments, failures and unresolved questions as part of an evolving investigation into document retrieval and factual evidence.
Explore potential collaborations →