Goals & non-goals

indx promises outcomes, not techniques. The routing technique, models, and parsers can change; the outcomes stay stable. The canonical list is core user features. The concrete feature-by-feature catalog, with examples and source links, is features.

The ambition behind these outcomes: the unstructured data inside an enterprise becomes AI-ready without anyone hand-wiring a pipeline per document type. Standard content stays cheap on CPU. Heavy models run only where they earn their cost. And the output — text, metadata, chunks, embeddings — stays yours, in an index you control.

Delivered by the current backbone

• An actionable plan. Submit content and business constraints. Receive a complete, explainable plan. The same source, constraints, policy, and snapshot always reproduce the same plan.
• Efficient execution. Execution respects declared quality, latency, cost, hardware, and residency constraints. Actual cost, latency, and fallbacks are recorded. Unsatisfied constraints are reported explicitly, never hidden.
• Specialization without losing generic coverage. A recognized document may nominate a specialist parser. The generic ladder always stays behind it as a fallback.
• A provable routing advantage. Planned and actual outcomes are compared against a fixed dataset and a heavy-pipeline baseline. See benchmarks.

Planned

These are target requirements. They wait until the backbone has measured baselines. The order they are built in is on the roadmap.

• Planned Classify a corpus on CPU. Standard inputs classified and organized without GPU processing. CPU-only rate and total cost reported at corpus scale.
• Planned Read difficult enterprise data. An eligible reading path for handwriting, drawings, spreadsheets, and large files. Today, handwriting is the benchmark’s one route miss.
• Planned Agent-readable export. Processed content exported to an index you control, with the embedding space and query encoder identified.
• Planned Deferred architecture. Region optimization, broad modality coverage, RAG export, persistence, and separate router/executor deployments. These follow the first measured slice.

Non-goals

These are refusals, not delays.

• indx does not own your index. Blocks and vectors are produced for an index you control.
• Planning never invokes expensive processing. No OCR, no model, no rendering, no network before the plan exists.
• Execution never silently re-plans. It may only use the routes and fallbacks the plan already named. A source that changed since planning is rejected with 409.
• No hook framework. Extension is keyed discovery with a reproducible snapshot, not ordered plugin invocation.

One consequence is easy to misread: manual review is a feature, not an apology. The ladder ends at a person so that no page is silently dropped. The FAQ explains why.