Ontology-Augmented Generation.

We use Ontology-Augmented Generation (OAG) for a proposed workflow that combines typed records, source passages, temporal information, and explicit review. It is a TACITUS research term, not an established external standard or a proven replacement for retrieval-augmented generation.

The ontology supplies a vocabulary: an actor is distinct from a claim, an allegation from an observation, and a commitment from a prediction. Context assembly selects the records and passages relevant to a question. The model then works with that package, while a reviewer can inspect what was included and what is missing.

The same design can sit on top of lexical search, vector search, a graph index, or a long-context model. Its value would have to come from better evidence selection, more useful review, or lower total cost. Valid JSON and a connected graph are insufficient evidence of any of those outcomes.

GraphRAG's official query documentation makes a related distinction: local search combines graph records with source chunks, while global search uses community reports for corpus-wide questions. A structured approach should be chosen for its fit to a task, not because a graph is automatically a better representation.

1. 01 · Retrieve

   Find candidate sources using text, metadata, dates, and permissions. Preserve their original versions.

2. 02 · Connect

   Propose entities and relationships. Keep ambiguous identities and unsupported links unresolved.

3. 03 · Assemble

   Choose records, source excerpts, counterevidence, and missing-information notices within a budget.

4. 04 · Review

   Check the interpretation and citations. Save the context version used for the answer.

This is where Context Capsules fit. A capsule is intended to carry a reusable selection of evidence and analytical context; it is not a dump of the entire graph. PRAXIS is the workspace for examining that material, asking follow-up questions, and developing a draft. Automated assembly across the full research pipeline remains under development.

Consider three fictional messages. Sam assigns ownership of a launch deck. Alex agrees to look at the figures, then later denies ownership of the whole deck. The question is whether the Thursday delivery was agreed. The source does not justify silently converting that exchange into an accepted commitment.

{
  "example_only": true,
  "question": "Was a Thursday delivery agreed?",
  "records": [
    { "id": "c1", "kind": "Claim", "speaker": "Sam",
      "text": "Alex owns the launch deck by Thursday.",
      "source": "msg-1", "recorded_at": "2026-04-13" },
    { "id": "c2", "kind": "Claim", "speaker": "Alex",
      "text": "I can look at the figures.",
      "source": "msg-2", "recorded_at": "2026-04-13" },
    { "id": "c3", "kind": "Claim", "speaker": "Alex",
      "text": "I did not agree to own the whole deck.",
      "source": "msg-3", "recorded_at": "2026-04-16" }
  ],
  "review_question": "Does msg-2 accept ownership or only a narrower task?",
  "candidate_commitment": { "status": "scope_disputed", "review": "required" }
}

A prose-only system could produce the same careful answer. The proposed benefit of the structured record is that a second task can reuse the distinction, inspect its sources, and revise it when new evidence arrives. The JSON above is explanatory, not a complete schema-conformant artifact or a measured model output.

Structure can remove repeated wording, resolve repeated references, and retrieve a small connected part of a case. It can also add identifiers, schema definitions, and relationship labels that make a prompt longer. We should measure token counts after serialization with the actual model tokenizer, not infer savings from the number of graph nodes.

Compression is a separate design choice. LLMLingua-2 (2024) learns which tokens to retain. SARA (ACL 2026) combines retained text with learned compressed representations. Those mechanisms differ from extracting claims into JSON; their results cannot be transferred to capsules by analogy.

The March 2026 Graph-RAG reasoning study reports configuration-dependent effects from graph-walk compression. We take that as a reason to test routing, model size, and evidence retention together. A short context that drops a negation, a deadline exception, or the opposing account is a failed context.

Li et al. (2024) found that sufficiently resourced long-context models outperformed retrieval on average in their tested setup, while retrieval retained cost advantages. This is an argument for a serious long-context baseline and hybrid routing, not a universal winner.

Our proposed test holds the source corpus, questions, model, answer requirements, and reviewer rubric fixed. Compare direct long-context reading, lexical/vector retrieval with reranking, a graph-based baseline, and reviewed capsule assembly. Run both equal-input-budget and equal-total-cost comparisons; one does not substitute for the other.

• Measure whether all answer-critical evidence reached the model, including counterevidence and relevant exceptions.
• Score answer correctness, citation support, uncertainty, and omission separately from schema validity.
• Remove typing, time fields, review, and graph links one at a time to identify which parts contribute.
• Repeat after a source correction and measure stale answers, update cost, and the time a reviewer needs to find the change.

The hypothesis fails if the extra structure produces no useful quality or review benefit at an acceptable cost. We have not published comparative OAG results. The TCGC evaluation plan explains the evidence we want to collect.

Primary sources reviewed September 29, 2026. Findings belong to their cited studies; they are not TACITUS performance claims.

1. From Local to Global: A Graph RAG Approach to Query-Focused Summarization

   Edge et al. · research paper · 2024

   Studies graph community summaries for questions about an entire corpus. Results concern the reported datasets and evaluation setup, not every retrieval or policy task.

2. GraphRAG query engine

   Microsoft · official documentation · Reviewed September 29, 2026

   Distinguishes local, global, DRIFT, and basic search; local search combines graph records and original text. Implementation documentation is not an independent performance comparison.

3. Retrieval Augmented Generation or Long-Context LLMs? A Comprehensive Study and Hybrid Approach

   Li et al. · EMNLP Industry Track · 2024

   Finds stronger average long-context performance when sufficiently resourced in its tested setup, alongside cost advantages for retrieval and a hybrid routing approach. The tested models and workloads do not determine the best choice for current deployments.

4. LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression

   Pan et al. · Findings of ACL · 2024

   Treats extractive prompt compression as token classification learned from distilled examples. Compression quality remains task-sensitive; preserving selected tokens does not prove preservation of every relevant relationship.

5. SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression

   Jin et al. · ACL long paper · July 2026

   Combines retained natural-language passages with learned compressed representations and evidence reranking under token budgets. The approach includes model training and mixed representations; it is not a drop-in JSON compression result.

6. The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA

   Zarrinkia et al. · preprint · March 18, 2026 · version 2

   Studies structured prompting and graph-walk compression on KET-RAG and LightRAG across three multi-hop QA benchmarks. Compression helps some configurations and hurts others. The results are not independently reproduced by TACITUS.