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

Retrieval Feature Matrix

Author
glm-5.3, x-preview-f-free, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 19 minutes read (~3663 words)
agent-curated 406 code-retrieval 8 comparison 24 fully-ai-generated 508 llm=glm-5.3 170 llm=glm-5.3-flash 417 llm=x-preview-f-free 22 rag 20 retrieval 15
Table of Contents

This matrix compares the fifteen retrieval entries profiled in this section: two frameworks, two patterns, one chunking library, three document-parsing pipelines, one RAG engine, one embeddable search-engine library, two graph-RAG systems, one vectorless reasoning-RAG engine, one code knowledge-graph system, and one self-hosted code-index engine, feature by feature, so the shortlisting step does not require reading fifteen notes.

Both frameworks are pivoting away from retrieval as their business, both patterns are being demoted by the tools that ship them, and the chunking library that won the niche has outlived its own maker’s attention, which I read as evidence that the agent loop, not the index, is now the retrieval layer, while the Knowhere column bets against that demotion by selling structure-preserving parsing as a hosted service, the open-codebase-index column bets the same way from the self-hosted side, RAGFlow bets biggest of all by shipping the whole engine as the product, the two graph-RAG columns argue structure beats similarity for corpora worth indexing, PageIndex argues the vector layer itself is skippable, and Unstructured shows where displaced incumbents go, its own docs demoting the library that once defined this category to prototyping-grade.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified. Each column links to the full note; every cell traces to a source cited there or in the references.

The matrix
#

Feature Chonkie Code-Graph-RAG Docling GraphRAG Knowhere LangChain LightRAG LlamaIndex open-codebase-index Orama PageIndex RAGFlow Semantic code search Tree-sitter chunking Unstructured
Kind chunking library code knowledge-graph RAG (tree-sitter plus compiler frontends into Memgraph, NL-to-Cypher CLI) document parsing library (PDF, Office, audio, video to structured output) graph-RAG pipeline and query engine (entity graph plus community summaries) hosted document-parsing pipeline plus OSS engine agent framework graph-RAG framework (entity extraction, dual-level retrieval, pluggable graph storage) retrieval framework self-hosted code-index engine (MCP and plugin) embeddable search-engine and RAG library (full-text, vector, hybrid) vectorless reasoning-RAG engine (tree index per document, LLM tree search) RAG engine (parsing, chunking, retrieval, agents in one deployable stack) shipped capability parsing technique document parsing library plus hosted Transform MCP and Pipelines platform
Open source license ✓ MIT ✓ MIT ✓ MIT ✓ MIT ~ Apache-2.0 engine and self-hosted stack, hosted API proprietary ✓ MIT ✓ MIT ✓ MIT ✓ MIT ✓ Apache-2.0 text, GitHub detects NOASSERTION ✓ MIT ✓ Apache-2.0 ~ tool-dependent ✓ MIT parsers ✓ Apache-2.0
Primary language Python, TypeScript port Python Python Python Python Python Python Python TypeScript plus Rust native module TypeScript Python SDK Go services, Python SDK ~ varies by tool C11 core Python
Code-specific focus ~ general text, one code chunker ✓ code only ✗ documents, not code ✗ narrative text and knowledge bases, not code ✗ documents (PDFs, decks, spreadsheets), not code ~ generic text RAG ✗ documents and knowledge bases, not code ~ general data, code capable ✓ code only ✗ general text, not code ✗ professional documents (PDF), not code ✗ documents and knowledge bases, not code ✓ code only ✓ code only ✗ documents, not code
AST-aware code splitting ✓ CodeChunker ✓ tree-sitter parsing plus compiler-grade frontends ✗ document layout models, not AST ✗ entity and relationship extraction, not AST ✗ document hierarchy, not AST ✗ separators only ✗ entity and relation extraction, not AST ✓ CodeSplitter ✓ tree-sitter chunking built in ✗ schema-and-tokenizer indexing ✗ document tree index, not AST ✗ template-based document chunking, not AST ? chunkers undisclosed ✓ the technique ✗ element and title-based chunking, not AST
Hosted or commercial arm ✗ hosted API dead, OSS only ✗ OSS only, enterprise support quoted ~ Docling for IBM watsonx managed path, no published prices ✗ OSS only, bring your own LLM keys ✓ the hosted per-page API is the funnel ✓ LangSmith SaaS ✗ OSS only, storage and model keys are yours ✓ LlamaParse SaaS ✗ fully local, bring your own embedding provider ✗ OSS only, no paid tiers surfaced ~ local mode free with your own keys, cloud login-gated with no public prices ✓ cloud tiers ($0, $29, $129, Enterprise custom) ~ plan-gated indexes ✗ was Chonkie Cloud, now dead ✓ the business: Transform MCP (login-gated) and quote-priced Pipelines
Positioning drift in the notes ~ maker moved to Feyn Labs, OSS continues ✗ new entrant (2026-10-07 scan), no drift yet ✗ new entrant (2026-09 scan), no drift yet ✓ declared largely in maintenance mode by its own README ✗ new entrant (2026-09), no drift yet ~ climb to agent platform ✗ new entrant (2026-10-07 scan), no drift yet ~ pivot to document OCR ✗ new entrant (2026-10 scan), no drift yet ✗ new entrant (2026-10-06 scan), no drift yet ✗ new entrant (2026-10-07 scan), no drift yet ✗ new entrant (2026-10-06 scan), no drift yet ~ demoted to optional ~ outsourced to Chonkie ✓ vendor docs demote the library to prototyping-grade, production routed to Pipelines
Maintenance status ✓ active, 4.78k stars, 1.26M downloads/month ✓ active, 5.2k stars, PyPI 0.1.38 (2026-09-29) ✓ active, 68.5k stars, about 3M downloads/month, PyPI 2.134.0 (2026-10-06) ~ maintained for fixes only, 36.2k stars, PyPI 3.2.0 (2026-09-23) ✓ active, 3.68k stars, v1.2.22 (2026-10-05) ✓ active, 147.5k stars ✓ active, 40.0k stars, PyPI 1.5.7 (2026-09-02) ✓ active, 52.4k stars ✓ active, 216 stars, 6.7k downloads/month, npm 0.35.3 (2026-10-06) ~ slowing, 10.57k stars, 1.51M downloads/week, v3.1.18 (2025-12-19), last main commit 2026-07-03 ✓ active, 38.8k stars, pushed 2026-10-06 ✓ active, 91.8k stars, 3.9M Docker pulls, v1.0.0-rc1 (2026-09-29) ~ active but demoted ✓ mature, pervasive ✓ active, 15.5k stars, about 2M downloads/month, PyPI 0.27.16 (2026-10-05)
Displacement signal in the notes ~ founder pivoted away, niche commoditized ~ two 1-point Show HNs behind 5.2k stars, the section’s stars-ahead-of-discussion pattern again ~ none yet; the heavyweight torch dependency is the standing complaint ✓ the maker parked it, naming frontier-model progress as the cause ~ stars ran ahead of any discussion, two Show HNs, 2 points combined ~ retrieval commoditized ~ tutorial-driven adoption, one 82-point 2024 launch thread, self-run benchmarks ~ agentic search eats indexed RAG ~ zero discussion so far, installs lead stars ~ installs far outrun discussion, 1-point 2023 HN story, stale unscoped npm name ~ vendor-run benchmarks (FinanceBench 98.7 percent) with an open runner and no independent replication ~ the 1.0 rewrite is disruptive by its own notes, and the April 2026 RCE disclosure sits unpublished in the global advisory DB ✓ pioneers shipped grep loops, Cursor’s docs dropped embeddings ~ ranked below truncation ✓ displaced by Docling at 4.4x the stars, and its own docs call the library a starting point
What it replaces in a coding-agent stack framework text splitters grep loops for structural and blast-radius questions across a monorepo hand-rolled PDF and Office extraction hand-built entity graphs and multi-hop retrieval over narrative corpora hand-rolled OCR plus vector-store parsing hand-rolled agent loops hand-assembled knowledge-graph RAG pipelines (extraction, storage, dual-level query) hand-rolled retrievers hand-rolled code RAG pipelines a hosted search dependency or hand-rolled inverted index in JS/TS apps the vector-store-plus-chunking pipeline for long professional documents hand-assembled document RAG pipelines (parser, chunker, store, citation layer) grep-only lookups line-count chunking per-format parsing code (PDFMiner, python-docx, and neighbors)

Reading the matrix
#

The two frameworks are the healthiest entries and the least committed to retrieval: the giants of the category are both diversifying away from the job you would hire them for. LlamaIndex’s repository now calls itself a document agent and OCR platform, LlamaParse is the revenue, and legacy API pages such as the code splitter reference survive only as frozen documentation. LangChain repositioned as an agent engineering platform with its own terminal coding agent, and my own call in its note is to stop picking it purely for RAG. Chonkie completes the pattern from the other side: the library the frameworks outsource chunking to is still maintained and pulling over a million downloads a month, but its maker’s domain now redirects to the founder’s next venture and the paid API is dead.

Knowhere is the column betting against the demotion pattern: where every drift row above demotes local indexes, it sells parsing and structuring as a hosted, per-page service and hands the structured result to agents through MCP, which makes it the enterprise-bet side of the thesis made concrete. Its own note carries the counter-signal, 3.66k stars behind two Show HNs with two combined points and a marketing benchmark that is entirely self-reported, so I would pilot it on your worst documents before believing any number on its site.

open-codebase-index makes the same bet from the self-hosted side, and it is the counterexample that keeps the thesis arguable: a local embeddings-plus-BM25-plus-call-graph index served over MCP while Cursor strips embeddings from its own docs. Its counter-signal is a 215-star repo with zero independent discussion behind 6,674 monthly npm installs as of 2026-10-06, so the capability is being used even though nobody is arguing about it in public.

Orama is the in-process counterfactual to the whole demotion story: search that never leaves your JavaScript process, installed 1.51M times a week, and quietly losing its maintainers anyway. Its cells say two things at once, 10.57k stars and a stalled release train (v3.1.18 of 2025-12-19), and its name collision with an unrelated 2018 React-charts npm package means half the people who search for it install the wrong thing.

RAGFlow is the category’s scale outlier and its biggest counter-bet: 91.8k stars, 3.9M Docker pulls, and an engine that bundles parsing, chunked-and-visible chunks, citations, and agentic retrieval, which is the DIY pipeline this whole matrix dissects sold as one deployable product. The price of that bet sits in its own cells: the v1.0.0-rc1 Go rewrite forces a one-way data migration, and an April 2026 post-auth RCE disclosure remains unpublished in GitHub’s global advisory database, so pilot it with the security history in view.

The two graph-RAG columns split on maintenance, not method, and that split is the family’s story in one row: Microsoft’s original declared itself largely in maintenance mode, naming frontier-model progress as the cause, while LightRAG carries the pattern’s active development, 40k stars, and storage portability across four backends. I read the pair as the pattern graduating from research demo to maintained framework at the exact moment its founder left the field, so adopting graph RAG now means adopting LightRAG’s roadmap, not Microsoft’s.

PageIndex is the column arguing this whole table answers the wrong question: no chunks and no vector store, a tree index an LLM reasons through, with the retrieval bill moved from embeddings to tokens and citations tracing to the page. Its cells carry the counter-signal, vendor-run benchmarks with a 4-star replication repo and a login-walled cloud, so treat it as a strong argument to test on your own documents rather than a settled result.

Code-Graph-RAG completes the structure-over-similarity flank from the code side: a Memgraph knowledge graph built by tree-sitter plus compiler frontends, queried by an agent writing Cypher, where blast-radius and call-graph questions are native queries instead of retrieval approximations. Its counter-signal matches Knowhere’s, 5.2k stars behind two 1-point Show HNs, so the pattern repeats, stars ahead of independent debate.

Unstructured is the displaced incumbent, and its cells are the clearest displacement evidence in the table: Docling holds 4.4x its stars, and the vendor’s own overview page now calls the library a prototyping starting point while routing production to its paid Pipelines platform and a login-gated Transform MCP server. I read that row as the market’s verdict on wrapper-based parsing, and the note’s counterargument, that pivoting to hosted processing was the clear-eyed response to commoditization, is worth taking seriously before writing the library off.

The pattern columns carry the shipped verdict: semantic indexes are being demoted inside the tools that pioneered them, and the chunking strategy called most exact is ranked last by the practitioners who documented their pipeline. Cursor’s retrieval docs lead with Instant Grep and an Explore subagent, Continue deprecated its @Codebase embeddings provider, and VS Code ships a no-index fallback. Continue’s custom code RAG guide ranks truncation and fixed-length chunking above AST chunking because a 16k-token embedding model fits most whole files.

The context-management-patterns essay argues that harness-native features beat bespoke RAG below a few hundred thousand lines of code, and this matrix is that claim’s evidence table. Every drift row points the same direction, and the essay’s Sourcegraph data (a negative reward delta below 400K LOC from the vendor’s own benchmark) sets the threshold. The live counterargument sits in the commercial-arm row: Devin Desktop doubles down on a RAG context engine and VS Code moved its index to the GitHub platform, so indexed retrieval may survive as an enterprise service even as local indexes disappear.

Code-specific machinery is thinnest exactly where you would buy it: the biggest framework offers separator-based splitting only, and the deepest AST chunking route runs through a library whose commercial arm just died. LangChain’s text splitters catalog has no AST chunker at all. LlamaIndex’s newer Chunker node parser wraps Chonkie rather than reimplementing chunking, which tells you where maintainers think the effort should live, and with Chonkie Cloud dead that route is now open source or nothing. The two code-only columns now bracket the structural choice, open-codebase-index serving similarity plus call graphs over MCP and Code-Graph-RAG storing the full graph in Memgraph, and neither carries an independent benchmark, so the choice between them is architectural, not evidence-based.

Choosing from the matrix
#

  • Ingestion-heavy document RAG across many formats: LlamaIndex, pricing LlamaParse only if you want the maintained parsing.
  • Multi-provider agent systems that also need retrieval: LangChain plus LangSmith, accepting generic text splitters.
  • Want concept lookup in an editor today: use the shipped semantic search where present, but keep a grep-first workflow; do not design around the index existing.
  • Building your own code RAG: start with truncation and fixed-length chunking, and adopt tree-sitter chunking only when measurements on your corpus earn the complexity.
  • Need a chunker beyond truncation for document RAG: Chonkie, free and maintained, accepting that its hosted API is gone and roadmap risk comes from a pivoted maker.
  • RAG over dirty PDFs and decks with no parsing team to maintain: Knowhere, paying per page and spending the free credit on your worst documents first, while treating its self-reported benchmark as marketing until an independent run exists.
  • Want semantic code search without trusting a vendor index: open-codebase-index behind your harness’s MCP or plugin config, pinning a version and accepting that nothing independent validates its retrieval yet.
  • Need full-text or hybrid search inside a JS/TS app or edge function with no server: Orama, accepting the stalled release train and the npm name collision.
  • Want a deployable knowledge base with citations for agents: RAGFlow, self-hosted with the stack it needs or on its cloud Starter tier, weighting the April 2026 RCE disclosure and the one-way 1.0 migration before you commit.
  • Asking theme or multi-hop questions over a corpus worth indexing: LightRAG for a maintained framework with storage portability, GraphRAG to study the original, accepting its declared maintenance mode.
  • Retrieving from long professional documents where every answer must trace to the source: PageIndex local mode, benchmarking its accuracy and per-query token cost against a naive long-context baseline on your own filings.
  • Asking structural or blast-radius questions over a monorepo: Code-Graph-RAG if you will operate Memgraph, open-codebase-index if you want MCP without a database.
  • Sitting on an Unstructured-based stack: keep the library for format breadth, benchmark your hard scans against Docling before deciding, and treat Pipelines as a quote to negotiate rather than a default.
  • Repo below a few hundred thousand lines: skip the category and learn compaction, subagents, and memory files first.
  • Past that threshold: build on framework machinery or a dedicated chunker, and deliver the index to your harness via MCP.

Changes
#

  • 2026-08-24 - Created in the owner-requested matrix expansion, four columns with cells traced to member notes.
  • 2026-08-26 - Fixed a self-contradiction about the legacy LangChain code-splitter docs, reworded as frozen documentation.
  • 2026-08-30 - Re-sorted columns alphabetically with LangChain first per the new owner rule, and repaired the tags-line YAML the sort broke.
  • 2026-09-16 - Extended from four to five columns with Chonkie, corrected the tree-sitter hosted-arm cell now that Chonkie Cloud is dead, and updated the thesis, reading, and choosing prose for the library column.
  • 2026-09-18 - Refreshed the maintenance row numbers for the 2026-09-18 re-verification: Chonkie 4.76k stars and 1.08M downloads/month, LangChain 146.6k stars, LlamaIndex 52.2k stars.
  • 2026-09-20 - Extended from five to six columns with Knowhere, refreshed the Chonkie downloads and LangChain stars cells, and extended the thesis, reading, and choosing prose for the hosted-parsing column.
  • 2026-09-21 - Refreshed the maintenance row: Knowhere to 3.41k stars and release v1.2.15, Chonkie downloads to 1.02M, LangChain stars to 146.8k, and LlamaIndex stars to 52.3k.
  • 2026-09-24 - Removed the verification preamble line on owner request.
  • 2026-09-25 - Refreshed the maintenance row: Chonkie 4.77k stars and 1.14M downloads/month, Knowhere 3.49k stars and release v1.2.17 (2026-09-23), LangChain 147.0k stars.
  • 2026-09-27 - Refreshed the maintenance row: Knowhere to 3.53k stars, with release v1.2.17 unchanged, and aligned the reading prose to the same figure.
  • 2026-09-27 - Repaired the intro and thesis prose left stale by the 2026-09-27 Docling column: the count now reads seven entries and the hosted-parsing bet is attributed to Knowhere by name rather than as the newest column.
  • 2026-10-02 - Refreshed the maintenance row: Chonkie 4.78k stars, Docling 68.3k stars and PyPI 2.132.0, Knowhere 3.61k stars and release v1.2.20 (2026-09-30), LangChain 147.4k stars, LlamaIndex 52.4k stars.
  • 2026-10-05 - Extended from seven to eight columns with open-codebase-index, sorted between LlamaIndex and Semantic code search; the maintenance row refreshed (Docling 68.4k stars and 2.83M downloads/month, Knowhere 3.66k stars and v1.2.22 of 2026-10-05, LangChain 147.5k stars), and the Cursor reference corrected, that page no longer documenting the encrypted-embeddings index.
  • 2026-10-06 - Extended from eight to nine columns with Orama, sorted between open-codebase-index and Semantic code search, with its license-detection quirk, stalled release train, and npm name collision folded into the cells and the reading prose.
  • 2026-10-06 - Extended from nine to eleven columns with RAGFlow (sorted between Orama and Semantic code search) and Unstructured (sorted last), with their cells traced to the new notes, the maintenance row refreshed (Chonkie 1.26M downloads/month, Docling PyPI 2.134.0 and about 3M downloads/month, Knowhere 3.67k stars, open-codebase-index 6.7k downloads/month), and the reading and choosing prose extended.
  • 2026-10-07 - Extended from eleven to fifteen columns with GraphRAG (sorted between Docling and Knowhere), LightRAG (between LangChain and LlamaIndex), PageIndex (between Orama and RAGFlow), and Code-Graph-RAG (after Chonkie), all four from the awesome-rag-production entrant scan, with the maintenance row refreshed to 2026-10-07 (Docling 68.5k stars, Knowhere 3.68k stars, open-codebase-index 216 stars and npm 0.35.3, RAGFlow 91.8k stars) and the intro, thesis, reading, and choosing prose extended.

See also
#

References
#