TL;DR
- 01A conversational interface can sit on a durable knowledge system; the interface does not determine the storage model.
- 02Structured context can help select evidence, but graph extraction and summaries introduce their own errors.
- 03Any token-efficiency claim must include its baseline, quality target, indexing cost, update cost, and query workload.
PRAXIS is the workspace. Context Capsules are the reusable records we want people and agents to work from. The engineering problem is selecting enough evidence for a question without confusing a compact view with the complete case.
Revision: Rewritten around context selection, retrieval alternatives, invalidation, and full lifecycle costs. Adds a scoped GraphRAG result and removes claims that chat interfaces cannot support durable knowledge. Added the scoped select the record the task actually needs example and its evidence boundary.
Separate the record from the working context
An analyst asks which earlier commitments constrain a new proposal. The complete case may contain years of documents. The model’s working context should contain the material relevant to this question: selected commitments, conditions, dates, disagreements, and source passages. The durable record should retain the rest so another question can select a different view.
A chatbot can use retrieval, databases, and versioned memory. A graph interface can still lose evidence or mislead its user. Our distinction is therefore architectural: PRAXIS is the place a person works, while DIALECTICA is the ingestion and review boundary we are developing for structured capsule records. The useful property is inspectable continuity, not the absence of a chat box.
Choose a retrieval strategy for the question
Lost in the Middle showed sensitivity to the position of relevant information in document QA and key-value tasks with the models it studied. It motivates position-aware testing; it does not prove that a particular current model cannot use long context. Likewise, using a graph does not establish superior reasoning. Compare strategies by question type, evidence coverage, and answer correctness. [1]
| Approach | Useful starting point | What can go wrong |
|---|---|---|
| Direct document context | A small, bounded file whose qualifications matter | Irrelevant text consumes attention; position can affect retrieval |
| Lexical or vector retrieval | A question answered by a few identifiable passages | Relevant conditions or contrasting sources may be omitted |
| Structured queries plus passages | Known actors, dates, commitments, or relationships | Extraction and entity-resolution errors constrain what can be found |
| Graph/community summaries | Broad questions about patterns across a collection | Summary loss, indexing cost, and stale derived context |
What a published token reduction does and does not mean
The GraphRAG paper reports that root-level community summaries required over 97% fewer context tokens than its map-reduce source-text summarization condition for two evaluated collections: podcast transcripts and news, roughly one million and 1.7 million tokens. This is a comparison of query context in that setup, with a quality tradeoff relative to other global methods. It is not a 97% reduction in all processing cost, nor a result for TACITUS. [2]
The graph index and community summaries must first be produced. A changing corpus may require extraction, clustering, and summary updates. Local fact questions can have much simpler retrieval needs. A global sensemaking result does not establish a universal advantage over vector retrieval, direct context, or a carefully constructed SQL query.
Account for the whole workload
If preprocessing costs P and the average saving per comparable query is S, a simple break-even estimate is P divided by S, only when S is positive. Real workloads also have updates and failures, so report the assumptions rather than turning this expression into a headline. A smaller retrieved package that triggers extra follow-up calls can cost more overall.
Compaction and structured notes can help an agent continue a long task. Anthropic describes these techniques alongside just-in-time retrieval, while discussing the risk of discarding important context. We treat that engineering guidance as a design input, not independent proof that our capsule representation saves tokens. [3]
total cost = initial extraction and indexing
+ source and schema update work
+ sum(retrieval + prompt input + generated output + retries)
+ storage and review effort
report separately:
input tokens / output tokens / cached-input tokens
first-query latency / repeated-query latency
evidence recall / answer quality / reviewer correctionsA compact capsule view needs an escape route
Consider a summary saying that a delegation “supports the agreement.” If the source said support was conditional on a review clause, dropping that condition changes the answer. The selected view should carry the qualification or direct access to the passage, and identify what was omitted. A reviewer must be able to distinguish “not retrieved” from “not in the source record.”
Derived summaries need dependency records. When a source is corrected, an actor alias changes, or a review decision is withdrawn, affected summaries must be invalidated or explicitly marked stale. PROV-O gives useful terms for derivation and revision; it does not perform invalidation for the application. [4]
Evaluate whether the structure earns its cost
Build a mixed question set: exact lookup, temporal comparison, competing accounts, and corpus-wide synthesis. Hold the source snapshot and model configuration fixed. Compare direct context, conventional retrieval, and structured retrieval with and without summaries. Score missing qualifications and unsupported claims as well as relevance, latency, tokens, and review time.
The intended benefit is context that supports a difficult question and survives a handover. PRAXIS capsule workflows provide the starting surface; complete portability, task-bound agents, and a dependable context-selection policy remain under development. We have no measured TACITUS token-efficiency result to report here.
Select the record the task actually needs
Modern chat, retrieval and professional workflow systems can preserve institutional context and support review. The proposed capsule value is narrower: a pinned record, the selected method and known omissions travel together into a recurring task. The briefing packet contains only available members of the 32-record assignment boundary, not the other seven dossiers. Compare this selection against direct documents and capable retrieval using equivalent sources, caveat retention and total setup/review cost. [5]
SOURCES
- [1]Liu et al. Lost in the Middle: How Language Models Use Long Contexts (TACL 2024).Controlled document QA and key-value retrieval experiments with the models studied; not a fixed limit for later models.
- [2]Edge et al. From Local to Global: A GraphRAG Approach to Query-Focused Summarization.Primary paper. Its global-query evaluation and context-token comparison are specific to the reported datasets and baselines.
- [3]Anthropic. Effective context engineering for AI agents (29 September 2025).Primary engineering guidance on retrieval, compaction, structured notes, and context tradeoffs; not an independent benchmark.
- [4]W3C. PROV-O: The PROV Ontology (Recommendation, 2013).Entities, activities, agents, derivation, revision, and invalidation.
- [5]Worked example / implementation boundaryAuthored website demonstration or reference explanation. Not independent analytical validation or production integration proof.
