PageIndex synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

From document submission to retained navigation

PageIndex is a document knowledge and context-engineering system with shipped indexing, storage, retrieval tools and answering orchestration. This account covers the repository artifact at commit af92d9658a6a9c0ae31171b43198db28951af2f4, with an applicability cutoff of 2026-10-15. It includes local drivers, alternative builders, framework adapters, Cloud clients and notebook consumers, but not the separately operated Cloud server. The evidence is static implementation, declared design and attributed performance reports from SRC-1, not observed execution or causal experiments. The boundary and Source register establish that scope. No additional source reads were made for this synthesis.

The ordinary local path starts when a caller submits a PDF. Code extracts page text with PyPDF2 and builds a navigation tree through Flash or classic indexing. Flash's initial layout extraction is model-free; its default full refinement and summaries are not. Managed submission refuses selected unreadable or unsuitable results and checks limited page-bound predicates before assigning an ID and saving the document. These checks are operational admission, not certification of the hierarchy or document claims. Direct builders can return products that managed submission would refuse, and Markdown/file-output builders do not automatically install their results in the managed document store (RT-RTE-local-submit, RT-CLM-layout-extraction, RT-RTE-alternate-builders).

Refinement separates proposal from decision. A model can propose subsection headings, but code filters candidates and compares the resulting decomposition against a fixed search-cost and estimated-read-token objective. Model proposals must pass normalized printed-title occurrence, range, ordering and duplicate checks. An optional external headings cache enters through a different branch: page-span/kind selection and own-title exclusion, followed by shared ownership and cost filtering. Missing printed matches alone do not reject cached headings. Thus there is no universal printed-heading guarantee. The cache's initial author and prior approval remain unknown (RT-RTE-tree-refine, MEM-OBJ-headings-cache).

The local store separates extracted page strings, the derived tree and discovery metadata. Tree IDs and spans address material for code; titles, summaries and document descriptions guide model interpretation and relevance choices. The stored tree omits full node text, which can be reconstructed from the separately retained pages. Summaries may be raw short-leaf text, model-produced reductions, or fallbacks. Parent calls consume opening text and child summaries rather than every underlying passage. Retaining these products makes later selective reading possible, but does not establish their semantic fidelity or retain a justification for every generated statement (MEM-OBJ-pages, MEM-OBJ-tree, MEM-OBJ-doc, EPI-OBJ-node-summary).

From a question to delivered evidence and answer

The representative shipped workflow asks a configured model about an already retained local PDF. The caller supplies the question, document target and optional history. PageIndex constructs managed guidance and target context, registers tools, and delegates repeated model/tool progression to the external Agents SDK. The answering model chooses semantic next steps; PageIndex code resolves and executes requested reads. This is one model-led loop with function handoffs, not a team of independently assigned agents. The workflow account traces that request through delivery, supported by RT-RTE-document-chat.

Guidance asks for structure-first reading above twenty pages and direct page access for smaller documents. It does not enforce that order. Local discovery is paged and time-ordered, not a semantic-search service; the model compares names and descriptions. Structure tools return a text-free outline, and page tools deliver requested retained pages as tool evidence. Response budgets can omit later requested pages and tell the model how to fetch them. These are character heuristics, not hard context-window guarantees: the first valid page can exceed the page budget, and an oversized outline leaf can escape partitioning (RT-BAP-managed-guidance, MEM-RTE-document-readback).

Context reduction occurs through selective reads and separate indexing, summary and vision-call inputs. Omitted pages remain stored and reloadable; only returned products cross isolated call boundaries. The inspected chat drivers do not establish automatic compaction of accumulated conversation history. Responses and Messages expose appendable tool transcripts for caller-managed reuse; ordinary Chat Completions does not provide the same tool history or an automatic durable session store. Provider prefix-cache annotations are not evidence of history eviction or summary memory (RT-RTE-protocol-history, MEM-RTE-document-readback).

Checks have concrete but bounded force. Invalid page requests and unavailable documents can produce guided errors; mixed page requests may produce partial success. A turn-limit failure is returned as failure guidance rather than automatically retried with a larger allowance. The model interprets tool recovery options, while the caller can narrow a question or retry. Successful delivery means returning the runner's final answer value, not a separate reviewer's acceptance of it. No expected-answer oracle or demonstrated post-generation semantic acceptance gate is established in this selected workflow (workflow verification and recovery, workflow delivery, RT-RTE-document-chat).

Grounding and citations likewise require separate judgments. Imported document statements are evidence, not externally verified truths. Generated summaries, descriptions and answers have indeterminate entailment without actual input/output pairs; unknown entailment alone does not establish novel conjecture. Citation helpers resolve recognized document/page/block references and prepare display data, but do not determine whether the neighboring proposition follows from that material. A source-only instruction is a delivered request, not demonstrated compliance (EPI-OBJ-answer-text, EPI-OBJ-doc-description, EPI-RTE-citation-resolution, EPI-CLM-cited-grounding).

Scope, maintenance and alternative consumers

Local managed chat injects a document allowlist into tool dispatch and removes model-supplied private arguments before doing so. That control covers its scoped tool paths, not all SDK calls, caller-added tools or framework hosts. Cloud own-model document targeting does not inherit the same structural allowlist. Default local tool registration is read-only; management is opt-in, while direct SDK deletion remains available. Neither registration nor remote tool preapproval establishes a shell/filesystem sandbox. Local storage also does not imply local inference: selected content can be sent to the configured model provider (RT-RTE-document-chat, RT-RTE-framework-tools, RT-RTE-cloud-tools).

Persistence uses separate atomic file replacements, not one transaction across the document. Marker files govern recognized availability; listing can reconstruct the manifest cache. Deletion removes the marker before best-effort cleanup and withdraws future managed reads. It does not revoke already delivered transcript strings, guarantee secure erasure, or provide an undo mechanism. Maintenance repairs availability metadata rather than adjudicating false source statements (MEM-OBJ-doc, MEM-RTE-store-maintenance).

Alternative routes preserve important distinctions. Classic indexing uses model title/page judgments, an exception-excluding score denominator and bounded correction/fallback; continuation can retain errors. Classic sanitization can remove legitimate matching source phrases before evaluation. Getter normalization can replace a full summary with a prefix summary and infer conventional endpoints, changing what a consumer sees. These controls cannot be treated as universal semantic validation (RT-RTE-alternate-builders, EPI-RTE-document-sanitization, EPI-RTE-tree-normalization).

The vision notebook implements a separate progression: obtain a tree, ask a model to select nodes, map their IDs and spans to retained page images, then supply those images to an independent answer call. Unlike request-led local tool reads, code automatically supplies selected evidence to that second model. Child summaries similarly reach later parent calls without a parent model requesting them. These are supported automatic delivery paths, not observed successful retrieval. Cloud clients also discover remote instructions and tool schemas and admit them into agent construction without a per-tool human approval stage, but the frozen source does not reveal the live catalog or server-side retrieval and storage policy (RT-RTE-vision-notebook, RT-RTE-tree-refine, RT-RTE-cloud-tools).

What improves, and what remains unestablished

The strongest supported improvement contribution is document-product revision: fixed cost predicates admit changed decompositions, and summary transformations produce reduced navigation material for later consumers. This is implemented capacity, not measured improvement in future answering. No comparison establishes that criticism of a consumed theory improved PageIndex's capacity for future action (RT-RTE-tree-refine, RT-BAP-cost-selection).

For the optimizer's formal cost account, theory-builder condition 1, localized content, is wired; condition 2, consequential consumption, is wired. Conditions 3, content-directed criticism with revision or changed reliance, and 4, retention of criticism's result for another round, are absent within the inspected optimization route. Iteration changes trees under unchanged formulae, not the formulae through criticism. These local absences are not repository-wide claims about inaccessible model processing. Reflection is not established by inspecting a document tree: the record does not demonstrate the required self-representation/update loop. Computational candidate selection is established, but autonomy as a theory builder is not established. Exercised self-improvement of PageIndex's own organization is also not demonstrated: retained trees can reach later consumers, but no subsequent operation causally depending on an improvement-directed organizational change is observed. These are independent conclusions, not a ladder or a single negative score (RT-RTE-tree-refine).

Likewise, document-fed summaries are not trace-fed learning merely because later calls consume them. Raw interaction transcripts are available without an established automatic trace-fed durable update. Unknown cache provenance, host retention and server updates prevent a system-wide learning negative. The model components expose configurable inference endpoints, not wired operation-time parameter updates in the inspected helpers; exact deployed versions and provider-internal learning remain unknown (MEM-OBJ-headings-cache, RT-RTE-protocol-history, RT-RTE-cloud-tools, RT-CMP-index-model, RT-CMP-summary-model, RT-CMP-chat-model, RT-CMP-vision-model).

For long, text-extractable documents, the supported architectural distinction is retained navigation plus selected evidence delivery rather than loading every page into each answer call. For scanned or graphical material, local extracted strings cannot be assumed equivalent to the source; the notebook's image path is distinct and Cloud OCR internals remain unavailable. For tightly scoped local QA, actual lookup controls exist, but they do not transfer unchanged to cloud or caller-owned frameworks. None of these scenario-relative findings ranks answer quality. README benchmark figures remain attributed reported operation, not independently verified performance for these paths (MEM-OBJ-pages, MEM-RTE-document-readback, RT-RTE-vision-notebook, EPI-CLM-benchmark-results).

Pinned end-to-end traces showing exact retrieved evidence and returned answers would change the operation assessment. Independent expected-answer comparisons, with controlled variation of retrieval or refinement, would address benefit and component attribution. A criticism-driven revision of the cost account, retained and consumed in another round, would alter the theory-builder findings. A traced update to PageIndex's own behavior-determining organization and causally dependent later operation would address exercised self-improvement. Server implementation and frozen catalogs would permit a separate code-grounded Cloud assessment. None is supplied by the current records.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Stale attribution remains in memory integration issue 1; it is not an open disagreement about cache admission.	SRC-1; MEM-OBJ-headings-cache, RT-RTE-tree-refine; report-verification limit.	The handed reports and their supported optimizer branches agree that cache candidates bypass the model occurrence check. The memory issue still attributes an overbroad claim to runtime.	Treating that issue as evidence that a runtime mechanism correction remains outstanding. This carries the reconciliation's unresolved attribution conflict without reviving the repaired guarantee.	Corrected integration wording aligned with the current runtime record.
The comparison inventory omits a separate direction/signal classification for heading-cache consumption.	SRC-1; MEM-OBJ-headings-cache, RT-RTE-tree-refine; profile-verification limit.	Known downstream selection by node span, heading kinds and own-title exclusion; partial comparison inventories.	Complete route-by-route direction and selector comparisons for optimizer cache consumption. Existing pull/push witnesses remain supported.	A record-grounded inventory entry separately classifying that consumption operation.
No observed run or causal comparison was admitted.	SRC-1; RT-RTE-document-chat, RT-RTE-tree-refine, EPI-CLM-benchmark-results.	Static source and attributed README results; no target tests, notebooks, provider calls or experiments executed.	Successful recovery, faithful answer dependence, measured reliability or benefit, benchmark verification, component effects and exercised self-improvement.	Inspectable pinned traces and independent outcome comparisons, with interventions appropriate to each attribution.
Cloud server and dependency implementations are outside the frozen evidence.	SRC-1; RT-RTE-cloud-tools, RT-RTE-framework-tools, RT-CMP-chat-model.	Shipped client orchestration, dynamic catalog admission and adapter grants; not live server policies, provider weights or dependency runtime behavior.	Remote storage/form/update/acceptance classifications, universal authorization, provider-wide learning negatives, and deployed version-specific guarantees.	Frozen server/catalog evidence, resolved model identities and relevant dependency or deployment evidence.
Initial cache production and external caller control remain unresolved.	SRC-1; MEM-OBJ-headings-cache, MEM-OBJ-doc, RT-RTE-protocol-history.	Cache loading/filtering, metadata serialization and returned raw transcript interfaces.	Initial human/model approval, complete lineage, actual history reuse, and a complete trace-learning or trace-source classification.	Provenance/admission records and consumer traces for the cache, metadata and caller-retained histories.
Structural checks and retained/generated labels do not establish semantic fidelity.	SRC-1; RT-RTE-local-submit, RT-RTE-alternate-builders, EPI-OBJ-node-summary, EPI-RTE-citation-resolution.	Limited span checks, branch-specific heading checks, model classic verification, summary completion and citation lookup.	Complete interval/hierarchy validity, universal printed occurrence, source truth, summary entailment, or certified answer support.	Actual source/output pairs with checks of the relevant semantic claims, complete structural predicates where required, and claim-to-evidence assessment.
Available evaluators, examples and deep extraction heuristics were not exhaustively audited.	SRC-1; boundary coverage, RT-RTE-alternate-builders, EPI-CLM-benchmark-results.	Functional coverage of shipped loops plus selected test contracts; not every test body, fixture/example byte or parser heuristic.	Exhaustive evaluator inventories, system-complete absence of acceptance or learning, and complete extraction-fidelity judgments.	Targeted inspection of the remaining material paths for a named conclusion; frozen benchmark runners/data for performance claims.
Storage and withdrawal guarantees are local and bounded.	SRC-1; MEM-OBJ-doc, MEM-RTE-store-maintenance.	Per-file replacement, marker-first deletion, best-effort directory cleanup and manifest repair.	Multi-file rollback, secure erasure, universal forgetting or semantic retraction of earlier answers.	Transaction/recovery and copy-revocation mechanisms with relevant failure and downstream-consumer evidence.