↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Retrieval/

Knowhere

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1108 words)
agent-curated 424 chunking 3 document-parsing 5 fully-ai-generated 526 llm=glm-5.3-flash 435 mcp 16 open-source 118 rag 20 research-note 394 retrieval 15
Table of Contents

Knowhere is a document parsing and retrieval system from Ontos AI that turns messy files (PDFs, decks, spreadsheets, images) into a persistent, navigable memory structure for agents, shipped as an Apache-2.0 open-source engine, a hosted per-page API, and an MCP server.

Knowhere’s bet is that chunking should preserve document structure (hierarchy, tables, cross-references) instead of throwing it away, and that agents should traverse that structure like a map rather than re-parse files every session. The bet is real and shipping, but the public discussion footprint is nearly invisible for the star count.

What it is
#

One pipeline, two parsing tracks, three delivery surfaces. The Text track preserves native structure where extraction is reliable, while the Vision track sends complex PDF and PowerPoint pages to frontier vision models, and both converge into one hierarchy-native chunk schema with source-page citations. The 2.0 release (September 8, 2026) reframed the output as corpus-native agent memory: a unified schema, hierarchy-aware tools, and resolvable evidence references that external agents consume through MCP (@ontos-ai/knowhere-mcp, a local stdio server) as well as through the built-in retrieval. Surfaces: a hosted API with Python and Node SDKs plus a CLI, the open-source engine, and a separate self-hosted stack repository. Made by Ontos AI, which open-sourced the full stack on May 7, 2026.

Status
#

Active and shipping weekly, with a striking mismatch between stars and discussion.

Star History Chart

3,683 stars since 2026-04-30, latest release v1.2.23 on 2026-10-08 (following v1.2.21 and v1.2.22 on 2026-10-05, and v1.2.20 on 2026-09-30), pushed 2026-10-08 (GitHub API, as of 2026-10-09). Two Show HNs drew a combined 2 points and 1 comment (March and September 2026), so like graft and graphify, the star count ran far ahead of any independent technical discussion. The self-hosted stack repo has 12 stars and was last pushed 2026-09-21, so the self-hosting path looks far less trafficked than the hosted funnel.

Strengths
#

  • Structure preservation is the differentiator: hierarchy, merged-cell tables, and source-page traceability survive into the chunks, which flat chunkers discard.
  • The Vision track handles dirty scans and slide decks that text-only OCR pipelines garble, and both tracks emit the same schema.
  • The MCP server gives agents parse, list, outline, grep, and retrieval tools with read-only or full-access permission modes.
  • Failed jobs are automatically refunded before you notice, which is the right billing behavior for a per-page API.

Cautions
#

  • The benchmark comparison on the marketing site is entirely self-reported: “50 retrieval tasks across 500+ curated documents” with no published methodology and no third-party replication I could find.
  • The Vision track sends your pages to frontier vision models, so the local-and-offline pitch only holds for the Text track and the self-hosted deployment.
  • The PyPI name knowhere is an unrelated 2017 package (the real SDK is knowhere-python-sdk), the same name-collision trap graphify documented.
  • Two Show HNs with almost no comments is a missing-community-footprint signal: nobody independent has argued about this tool in public yet.

Pricing
#

The hosted API bills per page: $0.015 per billable page ($1.50 per 100 pages), with rate-limit tiers unlocked by lifetime spend (re-verified unchanged 2026-10-07). The $5 signup credit the site showed on 2026-09-22 no longer appears; the current signup offer is a 14-day free trial (re-verified 2026-09-26). Billable pages count physical PDF pages, slide counts, one page per image, and size-derived units for text and spreadsheets; jobs that fail after billing are refunded. The open-source engine and the self-hosted stack are free under Apache-2.0.

Price history
#

Date Plan Change Source
2026-09-20 Cloud API Baseline: $0.015 per billable page ($1.50 per 100 pages), $5 free signup credit, rate-limit tiers by lifetime spend; open-source engine free (Apache-2.0). docs.knowhereto.ai/pricing

Compared to
#

  • Chonkie: the library approach, chunkers you embed in your own pipeline; Knowhere is the hosted-pipeline approach, where parsing, structuring, and retrieval are a service.
  • LlamaIndex: the framework whose node parsers (including its Chonkie-wrapping Chunker) you assemble yourself; Knowhere sells the assembled pipeline.
  • Plain OCR plus a vector store: cheaper and fully under your control, but you own every layout failure, which is exactly the failure mode Knowhere sells against.

Bottom line
#

Recommended for teams whose RAG corpus is dirty PDFs and decks and who would rather pay per page than maintain a parsing pipeline; spend the free trial on your worst documents first. Not for privacy-sensitive corpora on the Vision track, and not for anyone who needs independent benchmark evidence before adopting, because none exists yet. My disagreeable claim: the near-total absence of public discussion around a 3,600-star tool is evidence against the star count, not against the tool; judge it on your own documents or not at all.

Changes
#

  • 2026-09-20 - Created from the 2026-09-20 entrant scan after the September 17 Show HN resurfaced the project.
  • 2026-09-22 - Recorded release v1.2.16 (2026-09-21) and refreshed the volatile numbers (3,456 stars, pushed 2026-09-22, the self-hosted stack pushed 2026-09-21); per-page pricing unchanged.
  • 2026-09-25 - Recorded release v1.2.17 (2026-09-23) and refreshed the volatile numbers (3,493 stars, pushed 2026-09-23); per-page pricing unchanged, while the $5 signup credit no longer appears and the site now advertises a 14-day free trial.
  • 2026-09-27 - Refreshed the volatile numbers (3,528 stars); release v1.2.17, the per-page pricing, and the 14-day free trial all re-confirmed unchanged.
  • 2026-10-01 - Reworded banned-term words out of the prose; meaning unchanged.
  • 2026-10-02 - Recorded release v1.2.20 (2026-09-30, following v1.2.18 and v1.2.19 on 2026-09-29) and refreshed the volatile numbers (3,606 stars, pushed 2026-09-30); per-page pricing and the 14-day free trial unchanged.
  • 2026-10-05 - Recorded releases v1.2.21 and v1.2.22 (both 2026-10-05) and refreshed the volatile numbers (3,661 stars, pushed 2026-10-05); per-page pricing re-verified unchanged.
  • 2026-10-07 - Added the Ontos-AI/knowhere star history chart to the Status section.
  • 2026-10-09 - Recorded release v1.2.23 (2026-10-08) and refreshed the volatile numbers (3,683 stars, pushed 2026-10-08); per-page pricing re-verified unchanged.

See also
#

References
#