PageIndex epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is how PageIndex turns imported documents into retained navigation and later answers, and which checks license reliance on each product. Scope and exclusions: see boundary. Evidence identities and layers: see the Source register. All additional source reads below use SRC-1 at that same revision. No target execution, tests, live services, supplied execution traces or causal experiments establish operation or benefit.

The public warrant claims are EPI-CLM-tree-retrieval, EPI-CLM-benchmark-results and EPI-CLM-cited-grounding; extraction's narrower model-free claim is RT-CLM-layout-extraction. These are heterogeneous claims, not a general claim that storing text certifies knowledge.

Coverage

“Direct” denotes source inspected for this member; “supplied” denotes the runtime finding, not independent source inspection. Coverage is functional, not an exhaustive source audit.

Entry point or operation	Source path in SRC-1	Coverage and records, or limit
Local PDF submit and store admission	pageindex/local_api.py, pageindex/local_store.py	Direct extraction/check/save/getter branches and complete store; RT-RTE-local-submit, EPI-RTE-document-withdrawal.
Flash layout tree, fallback and bookmarks	pageindex/flash/main.py, pageindex/flash/api.py, pageindex/flash/embedded_toc.py	Direct main orchestration; supplied wrapper/refinement RT-RTE-tree-refine. Embedded-TOC function inventory searched; deep bookmark reconciliation, heading and layout predicates remain uninspected here. Prevents an exhaustive judgment of structural fidelity or every heuristic evaluator.
Expansion proposal/check/cost admission and optimizer CLI	pageindex/tree_optimize.py	Direct model proposal, cache candidate construction, shared ownership filter, expansion selection, CLI dispatch/writer; supplied merge/formula/scheduler RT-RTE-tree-refine, RT-OBJ-cost-policy. New-issue reporting is not a write veto in the CLI.
Summary, retitling, fallback and description	pageindex/utils.py	Direct scheduler leaf/parent/finish, normalization and mapping; searched summary/description helpers; EPI-OBJ-node-summary, EPI-OBJ-doc-description. Description helper details remain supplied through RT-RTE-local-submit; no measured fidelity.
Classic TOC extraction, sanitization, verification, repair/fallback	pageindex/page_index_classic.py	Direct sanitizer, appearance evaluator, verification denominator and meta-processor disposition; supplied remaining generation/repair RT-RTE-alternate-builders.
Markdown heading/text construction and optional model passes	pageindex/page_index_md.py, run_pageindex.py	Direct heading extraction and node-text slicing, summary short-text branch and CLI routing/writing; supplied thinning/tree construction RT-RTE-alternate-builders. Uninspected thinning details prevent complete semantic-preservation judgment.
Local registered tools: browse/get metadata/structure/pages/remove	pageindex/agent_tools.py	Direct registered operation map, page delivery, remover and managed/citation prompts; supplied discovery/structure dispatch RT-RTE-document-chat. Local semantic ranking explicitly unavailable in searched source; model compares metadata instead.
SDK tree/page/structure helpers and citation utilities	pageindex/client.py, pageindex/utils.py	Direct getter dispatch, cloud-only block/image/retrieval gates, citation resolver and tree normalizer; EPI-RTE-tree-normalization, EPI-RTE-citation-resolution.
Chat, Chat Completions, Responses, Messages and streams	pageindex/client.py, pageindex/local_chat.py, pageindex/chat_stream.py	Supplied RT-RTE-document-chat, RT-RTE-protocol-history. These are provisional implementation findings; not independently re-read end to end here. Prevents an independent whole-driver audit.
Cloud HTTP/MCP and exposed remote operations	pageindex/cloud_api.py, pageindex/mcp_bridge.py, pageindex/client.py, pageindex/agent_tools.py	Direct endpoint inventory and client getter/query gates; supplied RT-RTE-cloud-tools. Includes deprecated submit-query/get-retrieval, remote block/image reads, document/folder operations and managed chat. Server algorithms, current catalogs, OCR and acceptance policy unavailable; prevents characterizing remote warrant. Folder bookkeeping is not itself a truth check.
Adapter/hosted tool grants	pageindex/integrations/openai_agents.py, pageindex/integrations/anthropic_sdk.py, pageindex/integrations/claude_agent_sdk.py	Supplied RT-RTE-framework-tools. No independent host-loop or dependency audit.
Six notebooks, including separate image answering	cookbook/pageindex-vision-rag.ipynb, cookbook/	Supplied boundary coverage and RT-RTE-vision-notebook. Embedded outputs not admitted as observed runs; other notebook paths remain supplied coverage only.
Benchmarks, tests and examples	README.md, tests/, examples/, assets/	Direct README claims; supplied boundary inventory. Test bodies/fixtures, example results and image pixels not independently assessed here. External runners/data unavailable. Prevents a measured QA verdict or inference that examples establish route execution.

This assessment therefore does not assert a system-complete absence of evaluators, learning or acceptance. Provider parameters are separately scoped by RT-CMP-index-model, RT-CMP-summary-model, RT-CMP-chat-model and RT-CMP-vision-model: operation-time inference is wired, parameter updates are not wired in the inspected helpers, versions are not pinned, endpoints can resolve differently, and provider-internal learning remains uninspected.

Epistemic-object inventory

Fields owned by a declaring record are supplied there, not redefined here. Content is truth-apt when it asserts document content, location or an answer; a routing instruction is not itself an answer claim.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-page-text	Imported statements on numbered pages	See record	SRC-1 pageindex/local_api.py	Extraction is not source endorsement or OCR.
EPI-OBJ-tree-structure	Titles, hierarchy and page-location assertions	See record	SRC-1 pageindex/tree_optimize.py	Structural checks do not certify topic completeness.
EPI-OBJ-node-summary	Descriptions of section content; sometimes raw text	See record	SRC-1 pageindex/utils.py	Raw reuse and generated descriptions have different warrant.
EPI-OBJ-doc-description	One-sentence document characterization	See record	SRC-1 pageindex/local_api.py	Generation uses an already compressed tree.
EPI-OBJ-answer-text	Answer propositions over the user's question	See record	SRC-1 pageindex/local_chat.py	Exact propositions/entailment not determinable without responses.
EPI-OBJ-citation-entries	Asserted document/page/block references	See record	SRC-1 pageindex/client.py	Resolving an identifier does not establish support for its attached claim.
RT-OBJ-agent-instructions	None as answer content; behavior policy	See record	SRC-1 pageindex/agent_tools.py	Prompt delivery differs from compliance.
RT-OBJ-cost-policy	Formal search-cost account under its assumptions	See record	SRC-1 pageindex/tree_optimize.py	Not a measured answer-quality model.

Authority-route ledger

Ledger rows split functions even where the cited record is a valid container. All results are possible implementation results, not observed events. “See” delegates source-native mechanics to the cited record; the relation, license and mismatch columns are this member's assessment. No row treats retention as acceptance or lifecycle integration. For the inspected local paths, no evidence-consuming acceptance of an ampliative answer or summary against a named truth criterion was established; narrower structural admission is implemented.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-local-submit — import	content transformation	implemented	EPI-OBJ-page-text	truth-apt transformation: acquisition/import	PDF → extracted page statements	Parser; extraction domain, not truth adjudication	See route	See route	Text becomes later evidence	Source warrant unknown; document association/page order retained, extraction fidelity unmeasured	Makes extracted content available for save	Store writer; parser output; permissive; retained-document horizon	SRC-1 pageindex/local_api.py	none	none	Graphical content may be missing; no external source validation.
RT-RTE-local-submit — bounds	check/evidence production	implemented	EPI-OBJ-tree-structure	no content change	Nonblank extraction, Flash result class, supplied span endpoints	Code; input readability/refusal policy and limited page bounds	Before save; see route	Exception or continuation	Enforcing	Only the tested predicates; not complete interval validity, hierarchy or factual truth	Failed checks block managed persistence	Submission driver; exception/control flow; enforcing; current submit	SRC-1 pageindex/local_api.py	none	narrow-check	Does not check all endpoint conditions; direct builders differ.
RT-RTE-local-submit — admit	disposition/acceptance	implemented	EPI-OBJ-tree-structure	no content change	Save eligibility after managed checks	Code decides, caller initiates; see route	Before store write	See route	Permissive	Structural operational admission only, not acceptance of all claims as true	Allows new document ID to enter local retrieval	Store caller; successful check return; permissive; later document use	SRC-1 pageindex/local_api.py	none	admission-not-warrant	No per-content human approval; no QA criterion.
RT-RTE-tree-refine — propose	content transformation	implemented	EPI-OBJ-tree-structure	truth-apt transformation: indeterminate	Candidate subsection labels, locations and inferred grouping	Model proposes; cached headings alternate; see route	Full optimization above trigger	See route	Candidate only until later tests	Location/title hypotheses; printed titles may be copied, hierarchy entailment not established	Supplies candidates to filters	Optimizer; returned JSON/cache; advisory; candidate round	SRC-1 pageindex/tree_optimize.py	none	none	No concrete output to distinguish preservation from conjecture.
RT-RTE-tree-refine — model headings	check/evidence production	implemented	EPI-OBJ-tree-structure	no content change	Model-proposed heading/page correspondence	Code in propose_children: nonempty title, integer/range, normalized substring, duplicate/own-title and page-order checks	Model branch before shared ownership filtering/scoring	Keep/drop model-proposed headings	Enforcing within model branch	Normalized text occurrence on the proposed page and tested range/order predicates only; not relevance, completeness or heading semantics	Only passing model proposals reach shared filters/scorer	Optimizer; model candidate list; enforcing; current expansion	SRC-1 pageindex/tree_optimize.py	none	branch-limited-check	Cache candidates bypass this evaluator; matching source text can still be an inappropriate semantic heading.
RT-RTE-tree-refine — cache candidates	operational admission/selection/consumption	implemented	EPI-OBJ-tree-structure	no content change	Cached headings eligible for a node	children_from_cache selects node page span and configured kinds and excludes own normalized title	Cache branch before shared ownership filtering/scoring	Candidate heading/page list	Enforcing selection, not source-occurrence validation	Eligibility under supplied cache fields only; no independently established printed occurrence	Selected cached candidates reach shared filters/scorer without propose_children	Optimizer; cache-derived candidate list; permissive; current expansion	SRC-1 pageindex/tree_optimize.py	none	cache-occurrence-unknown	Cache provenance/prior validation uninspected; no observed unprinted candidate or acceptance.
RT-RTE-tree-refine — shared ownership	check/evidence production	implemented	EPI-OBJ-tree-structure	no content change	Known/ancestor duplication and shared-page ownership of model or cache headings	own_children rejects represented headings and headings outside ownership boundaries when printed matches establish position	Both branches before scoring	Keep/drop candidate headings	Enforcing tested predicates	Bounded ownership/duplication license only; a missing printed match alone decides nothing	Surviving candidates from either branch reach common span assignment and scorer	Optimizer; filtered candidate lists; enforcing; current expansion	SRC-1 pageindex/tree_optimize.py	none	narrow-check	Does not grant universal printed-heading occurrence or semantic-fidelity assurance.
RT-RTE-tree-refine — cost	check/evidence production	implemented	RT-OBJ-cost-policy	truth-apt transformation: entailed derivation	Candidate costs and token-read estimates under fixed formulae	Deterministic arithmetic over tree ranges/token counts; see record	Each expansion candidate	See route	Supplies ranking and predicate	Numerical comparisons within encoded proxy assumptions only; source ranges not thereby true	Produces scores used by selector	Selector; computed scalars; ranking; current optimization	SRC-1 pageindex/tree_optimize.py	none	proxy-not-outcome	Warrant is implementation deduction, not observed retrieval latency or accuracy.
RT-RTE-tree-refine — attach	disposition/acceptance	implemented	EPI-OBJ-tree-structure	no content change	Whether best decomposition replaces collapsed node	Code decides/vetoes against cost, gain-ratio and read-token predicates; see route	After filtering/scoring	Attach or remain collapsed	Enforcing	Acceptance for encoded navigation-cost objective and current tree, not answer truth or universal source occurrence; model/cache eligibility differs	Changes next tree traversal and summary inputs	Tree/scheduler; in-place children; enforcing; remaining indexing and later saved reads	SRC-1 pageindex/tree_optimize.py	none	proxy-not-outcome	CLI reports new validation issues but still writes; not a general validity veto.
RT-RTE-tree-refine — summaries	content transformation	implemented	EPI-OBJ-node-summary	truth-apt transformation: indeterminate	Leaf text or opening text plus child summaries → description/title	Model; parsing/nonempty/fallback logic, not semantic evaluator	Settled nodes, children before parent	See object	Retained routing content	Generated text is a candidate characterization; no established entailment or non-entailment	Feeds parent summary and later retrieval prompts	Scheduler/answering model; summary strings; advisory; parents and later sessions	SRC-1 pageindex/utils.py	none	generation-not-acceptance	Short leaves are separately warranted raw reuse; generated and fallback branches differ.
RT-RTE-tree-refine — raw reuse	content transformation	implemented	EPI-OBJ-node-summary	truth-apt transformation: non-ampliative reshaping	Short leaf text → summary field	Code below token threshold	Leaf summary branch	Stripped source text	Permissive	Inherits only extracted-text warrant	Allows model-free summary value	Parent summary/structure reader; text field; advisory; indexing/later reads	SRC-1 pageindex/utils.py	none	none	Does not certify original document statements.
RT-RTE-tree-refine — completion	check/evidence production	implemented	EPI-OBJ-node-summary	no content change	Settled-tree stability and whether requested model calls all failed	Scheduler code, runtime consistency and call-success domain	Finish; see route	Raise or return tree with summaries/fallbacks	Enforcing	Successful generation/stability only; not summary truth or coverage	Prevents returning globally failed/inconsistent summary run	Flash caller; exceptions/tree; enforcing; current indexing	SRC-1 pageindex/utils.py	none	completion-not-truth	Individual failed summaries can become fallback text.
EPI-RTE-document-sanitization	content transformation	implemented	EPI-OBJ-page-text	truth-apt transformation: non-ampliative reshaping	Raw extracted evidence → redacted/framed model input	Regex substitution and delimiter escaping; see record	Classic prompt construction	Filtered document text	Alters model-visible evidence, not stored page truth	Warrant degraded for literal quotation/content involving matched phrases; no semantic safety guarantee	Removes selected strings from evidence shown to indexing/check models	Classic model; prompt text; enforcing substitution plus advisory framing; each call	SRC-1 pageindex/page_index_classic.py	none	redaction-not-truth	Phrase matching can remove legitimate document content.
RT-RTE-alternate-builders — classic check	check/evidence production	implemented	EPI-OBJ-tree-structure	no content change	Title appears or starts on proposed page	Model fuzzy judgment over sanitized page text; code aggregates nonexception results	After generated TOC/physical-index checks	Fraction and incorrect-entry list	Evidence to caller, not independent oracle	Bounded model assessment of location; no independent answer reference	Supplies repair/fallback decision	Meta-processor; fraction/list; ranking; current build	SRC-1 pageindex/page_index_classic.py	none	denominator-limit	Exceptions excluded; sanitized evidence differs from raw source.
RT-RTE-alternate-builders — classic disposition	disposition/acceptance	implemented	EPI-OBJ-tree-structure	no content change	TOC usable for builder continuation	Code thresholds plus model repair; see route	After check	Continue, bounded correction, fallback or fail	Enforcing	Operational TOC acceptance for building a navigation tree; not complete document truth	Returned repaired TOC may still contain errors	Builder; control flow/returned TOC; permissive; current and persisted product	SRC-1 pageindex/page_index_classic.py	none	repair-not-certification	Three correction attempts do not establish all entries correct.
RT-RTE-alternate-builders — Markdown	content transformation	implemented	EPI-OBJ-tree-structure	truth-apt transformation: non-ampliative reshaping	Recognized headings and their literal line segments	Regex/line slicing in inspected unthinned extraction branch	Markdown build before optional passes	Literal title/text nodes	Navigation input	Warrant limited to recognized syntax and copied text, not general Markdown semantics	Seeds hierarchy and optional summaries	Markdown builder; extracted nodes; permissive; generated file	SRC-1 pageindex/page_index_md.py	none	branch-limited	Optional model summaries indeterminate; thinning not independently assessed.
RT-RTE-local-submit — description	content transformation	implemented	EPI-OBJ-doc-description	truth-apt transformation: indeterminate	Cleaned tree → one-sentence document description	Model; see route	Before metadata save	See object	Later selection cue	Candidate characterization, not accepted source proposition	Guides browse-time document choice	Answering model; metadata description; advisory; later invocations	SRC-1 pageindex/local_api.py	none	compression-lineage	No actual description to establish preservation or added claims.
RT-RTE-local-submit — retain	retention	implemented	EPI-OBJ-page-text, EPI-OBJ-tree-structure, EPI-OBJ-node-summary, EPI-OBJ-doc-description	no content change	Accepted-for-storage document product → disk	Store writer; see route	Successful submit	See route	Availability, not truth endorsement	Retains source associations and generated metadata; no new epistemic license	Makes pages/tree discoverable and reloadable	Local tools/getters; JSON files; permissive; until deletion	SRC-1 pageindex/local_store.py	none	retention-not-integration	No claim-level accepted ampliative product is established to undergo lifecycle integration.
EPI-RTE-tree-normalization	content transformation	implemented	EPI-OBJ-tree-structure, EPI-OBJ-node-summary	truth-apt transformation: non-ampliative reshaping	Wire aliases/range defaults → exposed tree	Deterministic assignment; see record	Getter return	See record	Changes exposed navigation	Copied descriptions keep prior warrant; default endpoints are compatibility conventions, not measurements	Exposes normalized values to readers	Getters/agents; returned JSON; permissive; current consumer	SRC-1 pageindex/utils.py	none	prefix-summary-loss	Both summary fields present: prefix overwrites summary. Earlier-server page endpoints are inferred, not checked.
RT-RTE-document-chat — select/read	operational admission/selection/consumption	implemented	EPI-OBJ-page-text, EPI-OBJ-tree-structure	no content change	Which documents/outline/pages reach answering context	Model semantic choices; local scope/budget code; see route	Tool calls within chat	See route	Scope enforcement plus advisory content	Imported evidence only; selection does not endorse truth	Allows selected evidence into next model call; budget can omit pages explicitly	Answering model; tool-result text; advisory, with enforcing scope; current chat	SRC-1 pageindex/agent_tools.py	EPI-CLM-tree-retrieval	no-relevance-oracle	First valid page can exceed budget; direct SDK reads have different filtering.
RT-RTE-document-chat — answer	content transformation	implemented	EPI-OBJ-answer-text	truth-apt transformation: indeterminate	Question/history/tool evidence → answer	Answer model, no demonstrated post-generation truth evaluator; see route	Final agent output	See route	Delivered answer	Candidate answer; source support and scope depend on actual output/evidence	Returns answer to caller without demonstrated semantic certification	Caller; answer text; advisory; current response, later history if supplied	SRC-1 pageindex/local_chat.py	EPI-CLM-cited-grounding	prompt-not-check	Neither entailed derivation nor ampliative conjecture can be established for unknown outputs.
EPI-RTE-citation-resolution	check/evidence production	implemented	EPI-OBJ-citation-entries	no content change	Citation tags → document/block lookup results	Code name resolution and optional server block lookup; see record	Explicit caller helper invocation	See record	Reference resolution only	Can associate a citation with accessible material, not establish claim support	Produces display data; no answer rejection/rewrite-for-truth	Renderer/caller; citation dictionaries; advisory; displayed answer	SRC-1 pageindex/client.py	EPI-CLM-tree-retrieval, EPI-CLM-cited-grounding	resolution-not-support	Unknown names and some missing blocks survive; Markdown citation format is not parsed.
EPI-RTE-document-withdrawal	lineage/freshness/recovery	implemented	EPI-OBJ-page-text, EPI-OBJ-tree-structure	no content change	Document availability after deletion or manifest repair	Marker/cache/schema checks, not truth evaluator	Management call or listing	See record	Enforcing future local lookup availability	Withdrawal from retrieval, not retraction of a false proposition	Blocks later store reads; cannot erase evidence already in caller history	Store/getters; marker and files; enforcing; future lookups	SRC-1 pageindex/local_store.py	none	deletion-not-retraction	No semantic invalidation propagated into earlier answers/transcripts.
RT-RTE-cloud-tools	operational admission/selection/consumption	implemented	RT-OBJ-agent-instructions, EPI-OBJ-page-text	truth-apt transformation: acquisition/import	Remote guidance/content → local model or caller	Remote server supplies; local client constructs; see route	Cloud API/MCP path	See route	Remote instructions and tools admitted	Remote evidence warrant unknown; client wiring licenses no server-internal truth assessment	Changes available tools, instructions and evidence	Agent/caller; MCP/HTTP; permissive/advisory; per-call/session	SRC-1 pageindex/agent_tools.py	none	opaque-server	Remote acquisition/evaluation/acceptance not determinable from this artifact.
RT-RTE-vision-notebook	operational admission/selection/consumption	implemented	EPI-OBJ-tree-structure, EPI-OBJ-answer-text	no content change	Model-selected nodes → page-image input selection	VLM relevance judgment; code maps ranges; see route	Before separate answer call	See route	Selection controls visible images	Selection is not checked relevance or answer warrant	Restricts evidence of independent answer frame	Answer VLM; selected images; permissive; second call	SRC-1 cookbook/pageindex-vision-rag.ipynb	EPI-CLM-tree-retrieval	no-independent-oracle	Generated answer relation remains indeterminate; cloud-produced tree opaque.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-tree-retrieval	See record	SRC-1 README.md; doctrine/design	Tree navigation and explicit references requested	RT-RTE-document-chat, RT-RTE-vision-notebook, EPI-RTE-citation-resolution	None supplied	No independent relevance/answer intervention; tools allow page access rather than forcing all reasoning through hierarchy	Implemented evidence-navigation and reference-resolution affordances	Traceable identifiers do not certify explanation or entailment; small documents may bypass structure by design.
EPI-CLM-benchmark-results	See record	SRC-1 README.md; reported operation	Attributed numerical results and external runner links	RT-RTE-document-chat is a plausible mechanism, not the frozen experiment	None admitted	Missing datasets, runner, answers and controlled comparisons prevent verification or component attribution	README reports results; architecture is compatible with selective reading	No transfer of headline accuracy to all local/cloud paths, models or documents.
EPI-CLM-cited-grounding	See record	SRC-1 pageindex/agent_tools.py; doctrine/design and prompt-delivery implementation	Explicit source-only and citation instructions	RT-RTE-document-chat, RT-BAP-managed-guidance, EPI-RTE-citation-resolution	None supplied	Prompt delivery is not compliance; resolver checks references, not support	Implemented request for grounded answers with optional page references	No implemented post-answer truth acceptance established in supplied driver account.
RT-CLM-layout-extraction	See record	SRC-1 pageindex/flash/api.py; implementation/design	Extraction separated from optional model passes	RT-RTE-local-submit, RT-RTE-tree-refine	None supplied	Default full pipeline contains expansion/summaries; layout heuristics can err	Narrow LLM-free extraction claim supported, not a model-free indexing claim	Semantic fidelity unmeasured.

Bounded conclusion

PageIndex acquires document statements rather than establishing their external truth. In the local route, extraction preserves document/page association but leaves source warrant and extraction fidelity unknown; scanned or graphical content is not equivalent to the stored text (EPI-OBJ-page-text). It retains page text separately from navigation and generated descriptions. Later retrieval gives a model selected evidence, not a certified proposition set (RT-RTE-local-submit, RT-RTE-document-chat).

There are real evaluators with narrow domains. The model-proposal branch tests normalized printed-title occurrence, range, order and duplicates. Cache candidates instead enter through node-span/kind selection and own-title exclusion, without that occurrence veto. Both branches undergo shared known/ancestor and conditional shared-page ownership filtering, span assignment and cost selection; missing printed matches alone do not reject cached candidates. Thus attachment grants no universal source-occurrence license. Arithmetic derives proxy costs from encoded inputs. Cost-based acceptance grants a decomposition operational authority over subsequent indexing and later navigation, for the intended use of reducing encoded search/read cost, not verified answer error. Classic title/page verification consumes evidence through model judgments and changes repair/fallback/continuation, but its exception-excluding denominator and bounded repair prevent interpreting acceptance as universal correctness (RT-RTE-tree-refine, RT-RTE-alternate-builders). These conclusions survive the strongest local counterexamples to a blanket “no checks” account.

Model-produced summaries, retitles, document descriptions and answers are truth-apt candidates. Their individual preservation, entailment or added conjecture is not determinable without actual input/output pairs. The short-leaf raw-text branch is a concrete non-ampliative counterexample to labeling all summaries novel model claims. Nonempty results, scheduler consistency, storage and answer return are not evidence-consuming semantic acceptance. No accepted ampliative claim followed by lifecycle integration is established in these inspected local paths; this is not an assertion about inaccessible Cloud server internals (EPI-OBJ-node-summary, EPI-OBJ-doc-description, EPI-OBJ-answer-text).

Content can change at the evidence-consumer boundary: tree normalization may prefer a prefix summary, and page delivery omits budget-exceeding later pages with explicit continuation guidance. Citation resolution links names/tags to material when available but does not reject unsupported propositions. Deletion withdraws future store availability, not already delivered answers. These mechanisms matter to lineage and reliance even though none is a truth oracle (EPI-RTE-tree-normalization, EPI-RTE-citation-resolution, EPI-RTE-document-withdrawal).

Behavioral authority differs by consumer. Optimizer predicates enforce decisions; retained descriptions advise model selection; managed source-only instructions request behavior per invocation. Cloud tools/prompts can change the agent's available behavior without any locally established truth-apt acceptance route (RT-BAP-cost-selection, RT-BAP-managed-guidance, RT-RTE-cloud-tools). This is capability/guidance admission, not evidence of parametric learning. Improved future capacity, reflection, theory-directed revision, autonomy and self-improvement do not follow from retrieval or cost reduction. The separate bounded theory findings remain those of RT-RTE-tree-refine; no additional positive or system-wide negative is inferred here. Missing run and experimental evidence prevents exercised-reliability and causal-benefit conclusions, including independent verification of EPI-CLM-benchmark-results.

Shared records

Operative objects

EPI-OBJ-page-text

Label: Page text

Source-native identity: local pages entries with page_index and markdown. Natural-language extracted statements plus symbolic page keys, stored in pages.json. Produced by local PyPDF2 extraction and later read as document evidence. Distinct from the process RT-RTE-local-submit; no supplied operative-object counterpart. Source warrant unknown: source/page association persists, but no external truth check or measured extraction fidelity is established. SRC-1: pageindex/local_api.py.

pages = [{"page_index": i + 1, "markdown": text} for i, text in enumerate(page_texts)] --- pageindex/local_api.py

return [_scrub_surrogates(page.extract_text() or "") for page in reader.pages] --- pageindex/local_api.py

EPI-OBJ-tree-structure

Label: Tree structure

Source-native identity: structure nodes with titles, node IDs, page spans and children. Mixed natural-language heading strings and symbolic hierarchy/ranges; in-memory dictionaries and local tree.json or CLI JSON. Claims are about document organization and location. Distinct output object of RT-RTE-tree-refine and RT-RTE-alternate-builders, not those processes themselves; no supplied object counterpart. Generation/check/admission mechanics: see those routes. Printed-heading checking applies to model-proposed children, not all candidates. Cache candidates bypass propose_children; node-span/kind selection and own-title exclusion precede shared ownership/cost filtering. The shared ownership filter rejects represented headings and conditionally misplaced printed headings, but does not reject solely for missing printed matches. Neither branch certifies semantic fidelity, and cache admission additionally lacks an independent occurrence guarantee. SRC-1: pageindex/tree_optimize.py, pageindex/page_index_md.py.

if normalize(title) not in normalize(pages[page - 1]): continue # the heading must be printed on that page --- pageindex/tree_optimize.py

The cache counterpath and shared filter are directly inspected implementation evidence, consistent with RT-RTE-tree-refine and MEM-OBJ-headings-cache. This object remains the output tree, not a duplicate cache object. No execution shows an unprinted heading being supplied or attached.

cached = own(children_from_cache(node, args.cache, args.kinds)) if cached: candidates.append(("cache", cached)) --- pageindex/tree_optimize.py

if kinds and heading.get("kind") not in kinds: continue if normalize(heading["title"]) == normalize(node["title"]): continue --- pageindex/tree_optimize.py

printed = found(page, lambda line: key in normalize(line)) if key else [] if (is_node(page, child["title"]) or page == start and top and printed and printed[-1] < top[0] or page == end and bottom and printed and printed[0] >= bottom[-1]): continue kept.append(child) --- pageindex/tree_optimize.py

A deterministic Markdown witness, deduced without execution: input # A\nBody\n## B\nTail produces heading metadata for A at line 1/level 1 and B at line 3/level 2; slicing yields # A\nBody and ## B\nTail. This is copied content and syntax-based grouping, not new claims about the world. It applies before optional thinning/model passes. SRC-1: pageindex/page_index_md.py.

header_pattern = r'^(#{1,6})\s+(.+)$' --- pageindex/page_index_md.py

node['text'] = '\n'.join(markdown_lines[start_line:end_line]).strip() --- pageindex/page_index_md.py

EPI-OBJ-node-summary

Label: Node summary

Source-native identity: node summary, Markdown prefix_summary, and same-page fused-node retitle. Natural-language values in the tree, eventually in tree JSON for retained paths. Distinct produced material from RT-RTE-tree-refine, which remains a valid container process, and RT-RTE-alternate-builders; no supplied object counterpart. Leaf summaries use selected text; parents receive opening text and child summaries. A generated characterization need not preserve all meaning, but no actual unsupported proposition is supplied to prove ampliation. Generated relation is indeterminate; raw short-text reuse is non-ampliative. Nonempty parsing and fallback are completion controls, not truth acceptance. SRC-1: pageindex/utils.py.

if count_tokens(text, model=self._model) < self._small_node_tokens: return text.strip() --- pageindex/utils.py

[{'title': c.get('title', ''), 'summary': c.get('summary', '')} for c in children], --- pageindex/utils.py

if retitle: written = parse_title(reply) if written: node['title'] = written return parse_summary(reply) --- pageindex/utils.py

if not node['summary']: node['summary'] = fallback_summary(node, "" if children else self._leaf_text(node)) --- pageindex/utils.py

if (self._asked and not self._answered) or not _any_summary(self.structure): --- pageindex/utils.py

EPI-OBJ-doc-description

Label: Document description

Source-native identity: generated description stored in document metadata. Natural-language one-sentence characterization held in a Python string, doc.json and manifest metadata. It can influence later document choice without certifying source truth. Distinct from EPI-OBJ-node-summary: whole-document discovery cue, different model input and consumer. Closest supplied record is producing route RT-RTE-local-submit, not a duplicate object. Flash passes a cleaned top-level tree to generation; no per-description semantic acceptance is established. SRC-1: pageindex/local_api.py; later interpretation is requested in pageindex/agent_tools.py.

description = generate_doc_description( create_clean_structure_for_description(structure, top_level_only=True), model=self._summary_model, ) --- pageindex/local_api.py

• After discovery: 1 match or 1 clearly best match → proceed to read and answer without asking. Multiple equally relevant → ask user to pick. --- pageindex/agent_tools.py

EPI-OBJ-answer-text

Label: Answer text

Source-native identity: own-model final answer, cloud answer or notebook VLM answer. Natural-language propositions in returned text/protocol envelopes; no automatic local durable answer store is established. Different producers and inputs remain those of RT-RTE-document-chat, RT-RTE-cloud-tools and RT-RTE-vision-notebook. This is their output class, not a duplicate route or a claim that all share one evaluator. Retained tool history is separately RT-RTE-protocol-history. No candidate instance was inspected; exact truth-apt propositions and their relation to evidence are not determinable. Prompt grounding is a requested restriction, not a truth check. SRC-1: pageindex/local_chat.py, pageindex/agent_tools.py; mechanism evidence supplied by the cited routes.

EPI-OBJ-citation-entries

Label: Citation entries

Source-native identity: parsed citations dictionaries and resolved display anchors. Symbolic document/page/block identifiers with optional natural-language block text, in caller-returned dictionaries; no independent durable store. Distinct from answer propositions and from RT-OBJ-agent-instructions: references are produced/processed output, instructions are policy. Closest supplied mechanism RT-RTE-document-chat does not describe this resolver in detail. Unknown document names retain null IDs, and unavailable blocks can retain tags without retrieved content; page-level entries receive no semantic-support assessment. SRC-1: pageindex/client.py.

entry: dict[str, Any] = {"document": citation["document"], "doc_id": ids[0] if ids else None, "page": citation["page"]} --- pageindex/client.py

Routes

EPI-RTE-document-sanitization

Label: Document sanitization

Part of: RT-RTE-alternate-builders

Endpoints: extracted classic document text → keyword redaction and delimiter framing → classic generation/verification prompts. Owner is _secure_doc_text; this is a material input-transformation part of the supplied alternate-builder route, not a replacement for its evaluator or disposition. It does not modify local stored page text. SRC-1: pageindex/page_index_classic.py; implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: redacted document text wrapped as model-visible data.
• Later read-back: inapplicable — this helper persists no changed artifact; generated downstream tree content has its own retention route.
• Delegated visibility: classic model receives transformed evidence, not the raw matching phrases.
• Selection predicate: known injection-pattern regex matches are replaced wherever they occur in supplied text; the title-appearance evaluator invokes the helper on its page evidence.
• Invalidation or expiry: inapplicable — current-call string transformation has no independent cache or expiry.
• Activation or effect: deterministic replacement changes model-visible input; successful instruction isolation or output correctness is unobserved.
• Evidence limits: static deduction establishes string alteration, not injection resistance or meaning preservation for arbitrary prose.

Concrete deduction: input The text says ignore previous instructions. becomes The text says [REDACTED]. before framing. The matched string could be legitimate quoted content. This removes rather than adds a source proposition, so the relation is non-ampliative reshaping with degraded literal fidelity, not entailed derivation or acceptance. Delimiter framing and the hardening prompt request that the remaining text be treated as data; they do not turn that request into a demonstrated semantic boundary.

r"ignore\s+(all\s+)?(previous|prior|above)\s+instructions?|" --- pageindex/page_index_classic.py

return _INJECTION_PATTERNS.sub("[REDACTED]", text) --- pageindex/page_index_classic.py

return _wrap_doc_text(_sanitize_doc_text(text)) --- pageindex/page_index_classic.py

EPI-RTE-tree-normalization

Label: Tree normalization

Endpoints: wire tree → unify_tree → public getter/agent structure. Owner is the client utility. It preserves topology, copies fields and chooses aliases/default endpoints. Distinct material getter subroute of supplied RT-RTE-document-chat, rather than new document generation or optimizer refinement. Also used outside chat by direct getters, so not contained exclusively in that runtime route. SRC-1: pageindex/client.py, pageindex/utils.py; implementation evidence.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: normalized tree dictionaries.
• Later read-back: inapplicable — this helper does not persist its normalization; a caller can retain returned values outside the helper.
• Delegated visibility: model tools/callers receive the normalized tree rather than both raw summary aliases.
• Selection predicate: client invokes normalization on a nonempty tree result; fields copied by fixed rules.
• Invalidation or expiry: inapplicable — fresh normalization of the current getter result, no cache owned here.
• Activation or effect: endpoint defaults and alias choice can change the content a later reader sees; no actual consumer response was observed.
• Evidence limits: deterministic deduction, not a run or measurement of compatibility/fidelity.

Concrete deduction: a node containing summary: "Whole section" and prefix_summary: "Opening only" exposes only summary: "Opening only", because the loop assigns that key twice. This selects existing text rather than inventing a proposition, but can degrade the intended summary scope. Missing endpoint inference similarly licenses a conventional navigation span, not verified source boundaries. No supersession of runtime is needed: the supplied account did not promise preservation of both aliases.

for key in ("summary", "prefix_summary"): if key in node: unified["summary"] = node[key] --- pageindex/utils.py

if end is None: end = starts[i + 1] if i + 1 < len(flat) else page_count --- pageindex/utils.py

EPI-RTE-citation-resolution

Label: Citation resolution

Endpoints: answer text → tag parsing/name lookup/optional cloud block lookup → reference entries and optionally numbered display links. Owner is PageIndexClient.get_citations / resolve_citations. Distinct from RT-RTE-document-chat: explicit post-answer caller utility, not a required agent acceptance gate. Source is SRC-1: pageindex/client.py; implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: reference dictionaries or answer text with numbered links and reference data.
• Later read-back: inapplicable — helper writes no retained state; caller owns any storage/display.
• Delegated visibility: caller/renderer receives looked-up block material when available; this is not automatically re-fed to the answering model.
• Selection predicate: explicit caller invocation, recognized tag syntax, supplied document IDs or library enumeration; ambiguous names raise.
• Invalidation or expiry: uninspected — helper performs current lookups but no observed source/document change or later caller use was supplied.
• Activation or effect: lookup/display wiring is implemented; changed reader reliance is unobserved.
• Evidence limits: identifier association and formatting are not entailment checks; Markdown prose references are explicitly not parsed.

Target/domain: tag identity and accessible document/block association, not the proposition next to the tag. Evaluator: deterministic parser/name resolver plus remote block availability. Unknown names and selected missing-block errors leave entries; no correctness veto, answer revision, rollback or claim acceptance occurs here. Caller chooses to resolve; code handles lookup failure. This is a check/evidence-production route, not product admission. Its display reshaping changes notation without granting additional warrant.

if entry["doc_id"]: try: entry.update(self.get_block(entry["doc_id"], block_id)) except PageIndexAPIError as exc: # Local raises carry no status; 429/5xx propagate. if exc.status_code not in (None, 403, 404): raise --- pageindex/client.py

EPI-RTE-document-withdrawal

Label: Document withdrawal

Endpoints: SDK delete or registered remove operation → marker/file/cache removal → future lookup absence. Related maintenance: listing/get-metadata reconstruct or use schema-valid metadata caches. Owner: DocStore; tool wrapper resolves names and reports per-document outcomes. Distinct management/availability route, not duplicate of creation in RT-RTE-local-submit. Management tool exposure is accounted for by RT-RTE-framework-tools. SRC-1: pageindex/local_store.py, pageindex/agent_tools.py; implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: store deletion boolean; SDK/tool response and per-name statuses.
• Later read-back: future store reads require the marker; cached metadata can reconstruct discovery only for directories whose marker exists.
• Delegated visibility: management caller receives statuses; already delivered evidence/history is not reached by this operation.
• Selection predicate: caller-selected ID or tool-resolved names; all supplied name elements checked before any deletion, at most ten names per tool call.
• Invalidation or expiry: marker removed first, directory removal attempted and manifest entry dropped; no time-based truth expiry or semantic claim invalidation.
• Activation or effect: future getter availability is changed by code; exercised deletion and downstream answer effects unobserved.
• Evidence limits: file cleanup is best effort and this is not semantic retraction of earlier answers or removal from external transcripts.

Caller or tool-using model initiates deletion; code can veto malformed inputs and resolve missing names. No additional per-document human approval occurs inside the remover. The trigger is an explicit management request, not a falsity judgment. No recovery/rollback restores removed content; re-submission is an external subsequent action. Cached metadata validity checks license use as metadata, not endorsement of the document's statements.

if doc_dir is None or not (doc_dir / "doc.json").is_file(): return None return _read_json(doc_dir / name) --- pageindex/local_store.py

(doc_dir / "doc.json").unlink() --- pageindex/local_store.py

if doc_dir.is_dir(): shutil.rmtree(doc_dir, ignore_errors=True) --- pageindex/local_store.py

Claims

EPI-CLM-tree-retrieval

Label: Tree retrieval

Claimed operation/warrant: hierarchical indexing and LLM reasoning replace vector retrieval and provide traceable, explainable, context-aware retrieval. Source: SRC-1 README.md, doctrine/design. Closest supplied claim RT-CLM-layout-extraction concerns model-free extraction, not retrieval quality or traceability; distinct identity. Route comparison must distinguish a mechanism and explicit references from warranted answer support.

1. Index: generate a tree-structure index for each document
2. Retrieve: agentically search that tree with LLM reasoning --- README.md

PageIndex is a vectorless, reasoning-based RAG engine that mirrors how humans read, delivering traceable, explainable, and context-aware retrieval, with no vector DBs or chunking.

--- README.md

EPI-CLM-benchmark-results

Label: Benchmark results

Claimed operation/warrant: local quickstart evaluation, comparative selective-read cost figures and FinanceBench accuracy. SRC-1 README.md, reported operation, not admitted observed-run/causal evidence. Different datasets/configurations remain distinct reported results; none is inferred from another. No supplied claim counterpart after comparison with RT-CLM-layout-extraction; that concerns architecture, not benchmark outcomes.

PageIndex-OSS-Benchmark measures exactly the setup in the quickstart above (PageIndexClient() in local mode, flash indexing, no OCR) on 62 lookup questions over 34 PDFs (1,945 pages) drawn from MMLongBench-Doc-V2. Every question's answer is a fact stated in running text, so a wrong answer is a retrieval or reading failure, not a reasoning one. --- README.md

On documents where both routes return the same answer, native PDF input costs 2.1× more at 52 pages and 16.6× more at 420 (gpt-5.6-sol, prompt caching excluded) — and at 805 pages the document no longer fits in the context window at all. --- README.md

PageIndex reached a state-of-the-art 98.7% accuracy on FinanceBench (financial document QA benchmark), vastly outperforming vector-based RAG. --- README.md

EPI-CLM-cited-grounding

Label: Cited grounding

Claimed design contract: source-only answers stating only read evidence and supported citations. SRC-1 pageindex/agent_tools.py, doctrine/design embodied in delivered prompt text. This is a material part of the instruction object, not a new copy of RT-OBJ-agent-instructions; different record kind because this record identifies what the policy demands, not the stored policy itself. Distinct from extraction claim RT-CLM-layout-extraction.

"markdown": """\ GROUNDING - Answer only from the user's PageIndex documents. Call get_page_content() and state only what was actually read there. - Never fill a gap from general knowledge. When the documents do not answer the question, say so. --- pageindex/agent_tools.py

• Cite only statements supported by tool outputs, as a bracketed reference: [{docName}, p. {pageNumber}] or [{docName}, p. {pageNumber}, block {blockId}]. Place immediately after the claim. --- pageindex/agent_tools.py