Should you try PageIndex?
Trial PageIndex if your team can evaluate document-tree retrieval against known answers. Keep your existing retrieval system until the trial shows whether PageIndex improves answers or source tracing on your documents. The supplied sandbox run establishes command outcomes, not that comparison.
PageIndex replaces a vector index with a hierarchical tree that a model searches. The commit-pinned README describes two steps: build a document tree, then search it with LLM reasoning. Layout supplies the initial structure; a model summarizes and refines it. Local indexing therefore still involves model configuration.
What the September 28 run establishes
Our run on September 28, 2026, installed commit 037a7db in 55 seconds. The full recorded SHA is 037a7dbacfb9a19f38b354ce60cee5094b3f854c. Installation brought in 106 packages with a recorded footprint of 280 MB. The harness recorded a successful build command in 6 seconds and a successful pytest command in 148 seconds. Installation, build and tests each returned zero and did not time out.
The stored parser summaries report 703 passed, 0 failed, 100 skipped and total 703. Those fields are retained without calculating a new total. The raw footer says 703 passed, 100 skipped, 4 warnings in 139.04s (0:02:19). Its 139.04-second pytest duration and the harness's 148-second command duration are separate observations. Exhaustive unique test coverage was not verified.
The supplied run contains no live-model retrieval evaluation, Cloud exercise, answer-quality comparison or workload cost measurement. The MrKeyoor testing methodology explains how command outcomes and parsed counts reach reviews. Its supplied November version provides provenance context but does not prove which procedure applied to this September run.
Local storage still needs a model boundary
The commit-pinned README describes local mode for text-based PDFs, local index storage and page-level citations. It assigns managed indexing and storage, OCR, image understanding and block-level citations to PageIndex Cloud.
Local document storage does not settle where model processing happens. The README describes models writing summaries and searching the tree. The v0.2.20 release notes further explain that page content flows through your process to your provider using your credentials. Before submitting restricted documents, check both the storage destination and the model endpoint.
The tested commit differs from v0.2.20
The v0.2.20 release page identifies release commit 619cbd8; the sandbox record identifies 037a7db. Release documentation can explain that version's interface, but this run cannot certify its complete behavior.
The release notes describe a citation-contract change introduced in v0.2.19. The earlier list-returning method was renamed from resolve_citations to get_citations, while resolve_citations returns an answer-and-citations object. Pin the intended version and verify the citation output your application consumes before upgrading. The supplied run does not establish that this migration was tested.
The dependency audit needs follow-up
Pip-audit recorded 1 advisory. The supplied result does not identify the affected dependency, severity or remedy. Its packages field does not establish scan coverage, and no scannedPackages value is supplied. Inspect the advisory before adoption; this count is not a security assessment.
When a broader framework fits
LlamaIndex provides data connectors and customizable indices, retrievers, query engines and reranking modules. Consider it when assembling those components around your data sources is the main task. Its README says the open framework remains available while the company's primary focus has shifted toward parsing and extraction. The supplied evidence does not establish a working PageIndex integration.
Haystack supports modular pipelines with control over retrieval, routing, memory and generation. Consider it when you need to define how those stages interact. That orchestration scope is broader than PageIndex's document-tree retrieval. A working PageIndex integration was not established by the supplied sources.
What the ratings mean
The assessed scores are editorial judgments about the supplied evidence, not benchmark results. The rubric is: 1 means a substantial observed obstacle; 2 means major preparation remains; 3 means usable with material unresolved limits; 4 means the evidence supports a trial with clear caveats; 5 requires strong evidence across the assessed dimension. Dimensions without sufficient evidence remain unassessed.
Setup ease is 4 because installation and build completed, while model configuration and a real document workload remain untested. Documentation is 4 because the supplied README and release notes explain storage boundaries, model roles and the citation-contract change. Maturity is 3 because successful command outcomes coexist with a documented API change, an unresolved advisory and no retrieval-quality evaluation from this run.
Community is unassessed. Published releases and integration documentation establish release communication, but the supplied evidence does not establish adoption, contributor diversity or support responsiveness. It receives no numeric rating.
Run a document trial before switching
Select representative documents and questions with known answers. Check whether every relevant page is reachable, whether each answer is correct and whether its citation supports it. Compare the results with your existing retrieval system, then measure model cost and latency for the same workload. Include scanned or image-rich documents only in a configuration whose documentation supports them.