We Compared The Features of 82 RAG Tools: Here's What We Found

Last updated: May 25, 2026

RAG tools look mature at the pipeline layer, but the features that make them operationally useful are still unevenly distributed. We built a dataset of 82 RAG tools, classified every feature with a seven-label availability scheme, and ran the aggregates to figure out what actually matters if you are shipping your own RAG tool.

The dataset spans seven workflow families: RAG application frameworks, managed RAG platforms, knowledge chat applications, retrieval and ranking components, evaluation and observability tools, memory and graph retrieval tools, and document ingestion and parsing tools. For each tool we recorded a comparable feature taxonomy across ingestion, retrieval, grounding, evaluation, memory, and governance, then classified availability based on actual packaging rather than marketing claims.

The full feature comparison table

We built this dataset from scratch. For each of the 82 RAG tools, we inspected public feature information and recorded the availability of 12 feature categories: document parsing and layout understanding, data connectors and source synchronization, chunking and semantic segmentation, embedding and vector index management, hybrid search and retrieval orchestration, reranking and relevance optimization, citations and source-grounded answers, agentic workflows and tool calling, knowledge graph and entity extraction, memory and personalization, RAG evaluation and quality monitoring, and deployment, governance, and enterprise controls. Each feature was classified with one of seven standardized availability labels. The full comparison table is below.

Questions on features of RAG tools

These are the questions we kept returning to while building the dataset. They matter if you are deciding which RAG tool features are non-negotiable, which ones can differentiate, which ones to gate, and what to build first.

Which features are commoditized in RAG tools?

The most commoditized features in RAG tools are deployment and governance, document parsing, citations, chunking, embeddings, and hybrid search. All six appear in at least 60 of the 82 tools, which makes them the core feature surface buyers expect to see.

Deployment, governance, and enterprise controls are the clearest category-wide capability, appearing in 68 tools. That does not mean every implementation is generous or complete, but it does mean the market now treats operational controls as part of the RAG product surface.

Document parsing and citations both appear in 65 tools, which is a strong signal that RAG tooling has moved beyond raw vector search. Buyers expect a product to ingest source material and preserve enough traceability to support grounded answers.

Chunking is almost as widespread, appearing in 64 tools. It is especially universal in RAG application frameworks, managed RAG platforms, and knowledge chat applications, where every tool in those three workflow families includes it.

Embedding and vector index management, plus hybrid search and retrieval orchestration, each appear in 60 tools. That pairing is important because it shows that RAG tools are no longer competing on vector search alone, but on the orchestration layer around retrieval.

The builder takeaway is simple: a general-purpose RAG tool that skips parsing, chunking, embeddings, hybrid retrieval, source-grounding, or deployment controls will look incomplete. Exceptions only make sense for narrow products like rerankers, parsers, or evaluation-only tools.

Which features are usually free by default in RAG tools?

In RAG tools, the features most likely to be free by default are embedding and vector index management, chunking, and hybrid search. Embeddings lead the category with 35 free-full cases among 60 present implementations.

Lower-level retrieval primitives carry the strongest free-full posture. Embeddings are free full in 58.3% of present implementations, and chunking is free full in 45.3%, which makes these the most open parts of the RAG stack.

This pattern is strongest in application frameworks and open technical components. LangChain, LlamaIndex, Haystack, RAGFlow, R2R, Kotaemon, and many smaller open-source tools expose core indexing or chunking functionality without a hard commercial gate.

Hybrid search is widely available but more often constrained than embeddings. Among tools with hybrid retrieval, 24 are free full and 28 are free limited, which suggests vendors are willing to expose retrieval orchestration but often cap scale, hosting, or integrations.

Document parsing is free in many tools, but usually not free full. Only 16 of 65 parsing implementations are free full, while 41 are free limited, which makes ingestion a more common freemium surface than a fully open one.

The rule for builders is to make the technical starting loop accessible: ingest a small corpus, chunk it, embed it, and retrieve from it. Charging too early for these primitives fights the category norm, but unlimited scale does not need to be free.

Which features are most often limited, paywalled, or premium-only in RAG tools?

The most gated features in RAG tools are deployment and governance, RAG evaluation, data connectors, and agentic workflows. Governance has the largest paid-only count with 20 paid-only implementations among 68 present cases, while agentic workflows have the largest restricted-access count.

Deployment and governance are both widespread and heavily monetized. Only 14 of 68 implementations are free full, while 20 are paid only, which makes this the strongest broad paywall in the dataset.

RAG evaluation is a smaller but cleaner premium signal. It appears in 44 tools, and 8 of those implementations are paid only, which reflects how monitoring, benchmarking, and quality assurance become valuable once a RAG system is in production.

Data connectors use a softer gate. Connectors appear in 44 tools, but only 9 are free full, while 24 are free limited and 5 are restricted. That means vendors often gate the number of sources, the sync frequency, the hosted connector layer, or the enterprise integrations.

Agentic workflows have the most unusual gating profile. They appear in 53 tools, but 16 are restricted, which is more than any other feature. In practice, tool calling is often available only in specific deployment modes, hosted products, integrations, or advanced workflow environments.

Paid-only and restricted gates also cluster by workflow. Memory and graph retrieval tools make deployment and governance paid only in 5 of 6 present cases, while document ingestion tools make it paid only in 6 of 17 present cases.

Knowing what to charge for is half the job. The other half is getting seen: our database of 450+ businesses breaks down how each one brings in customers.

Which features are still strong differentiators in RAG tools?

The strongest differentiators in RAG tools are knowledge graph and entity extraction, memory and personalization, RAG evaluation, and agentic workflows. They either sit below 65% penetration or have unusually fragmented access, which makes them more useful as positioning signals than basic pipeline features.

Knowledge graph and entity extraction is the cleanest differentiator because it appears in only 21 of 82 tools. When a product like Graphiti, Microsoft GraphRAG, TrustGraph, Cognee, or WhyHow.AI leans into graph retrieval, it is signaling a different architecture from standard vector-first RAG.

Memory and personalization is similarly differentiating, appearing in 23 tools. It is not heavily paywalled when present, but its scarcity means that products with persistent user, entity, or conversation memory can stand apart from generic document chat products.

RAG evaluation is a differentiator because it changes the buyer from a builder experimenting with retrieval to a team operating quality-sensitive systems. Evaluation appears in 44 tools overall, but it reaches 100% inside evaluation and observability tools.

Agentic workflows sit in a middle zone: common enough to matter, but restricted enough to differentiate. A product that makes tool calling broadly usable can stand out from products that mention agents but limit access to specific deployment or partner conditions.

The best differentiators are not random add-ons. They extend the RAG system from retrieval into reasoning over entities, persistent context, production quality, or workflow execution.

Which features are rarely offered in RAG tools?

The rarest features in RAG tools are knowledge graph and entity extraction, memory and personalization, and RAG evaluation when viewed outside specialist workflows. Knowledge graph features appear in only 25.6% of tools, and memory appears in only 28.0%.

Knowledge graph and entity extraction is rare because most RAG tools still treat retrieval as document or chunk retrieval. Only 21 tools include graph or entity extraction, and 8 of those sit in the memory and graph retrieval workflow.

Memory and personalization is also rare, with 23 present cases. Knowledge chat applications include it in 6 of 11 tools, but document ingestion, retrieval components, and evaluation tools largely avoid it.

RAG evaluation looks moderately common overall, appearing in 44 tools, but the distribution is highly uneven. It is universal in evaluation and observability tools, yet appears in only 4 of 11 knowledge chat applications and 3 of 9 retrieval and ranking components.

Data connectors are not rare at the category level, but they are rare in several technical workflows. Evaluation tools have zero connector coverage, and retrieval and ranking components include connectors in only 1 of 9 cases.

The important reading rule is that rarity in RAG tools is often workflow-driven. A feature can be rare across the full dataset and still be table stakes inside a narrow sub-category.

Which missing features create the biggest opportunity in RAG tools?

The biggest feature opportunities in RAG tools sit where established workflow boundaries leave obvious gaps: memory in managed RAG platforms, knowledge graphs in knowledge chat applications, and evaluation in document ingestion tools. These gaps are large enough to create differentiated products without requiring a completely new category.

Managed RAG platforms are strong across the core stack, but only 4 of 12 include memory and personalization. A managed platform that combines hosted ingestion, retrieval, evaluation, and persistent memory would occupy a more complete application infrastructure position.

Knowledge chat applications have full coverage across parsing, chunking, embeddings, hybrid search, citations, and deployment, but only 1 of 11 includes knowledge graph and entity extraction. That creates room for chat products that understand entities and relationships rather than only retrieving source chunks.

Document ingestion tools have universal parsing and strong chunking coverage, but only 8 of 23 include RAG evaluation. A parser that measures downstream retrieval quality could move from an ingestion utility into a quality-sensitive RAG infrastructure product.

Retrieval and ranking components have strong embeddings and hybrid search coverage, but no memory or graph coverage in this dataset. That absence creates an opening for retrieval components that optimize not only relevance, but also entity continuity and user context.

The opportunity pattern is not to add every missing feature everywhere. It is to find the adjacent feature that buyers already expect in the next workflow over, then bring it into a product where competitors still treat it as out of scope.

A sharp product still needs a way in front of buyers. See how 450+ businesses did it in our database, channel by channel.

What should be free versus paid in RAG tools?

In RAG tools, the free layer should cover the basic RAG loop: parsing a limited corpus, chunking it, embedding it, retrieving from it, and producing grounded answers. The paid layer should concentrate around scale, connectors, evaluation, governance, hosted operations, and advanced agentic workflows.

The basic RAG loop is where buyers expect hands-on exploration. Chunking, embeddings, and hybrid search all have large free-full or free-limited footprints, so hiding them completely behind a paywall makes a new tool harder to evaluate.

Parsing should be free enough to prove quality, but not unlimited. The dataset shows parsing as free limited in 41 of 65 present implementations, which makes usage caps, page caps, file-size limits, and hosted-processing limits normal for the category.

Connectors are safe to limit because they create operational value beyond the core algorithm. A small set of free connectors can reduce friction, while enterprise sources, sync controls, or managed connector infrastructure can sit behind paid plans.

Evaluation belongs closer to paid once the product moves from prototype to production. Since 8 of 44 evaluation implementations are paid only, the category already treats monitoring, quality gates, and advanced benchmarking as monetizable.

Governance is the clearest paid surface. Deployment, governance, and enterprise controls have 20 paid-only implementations, so buyers already expect security, admin, compliance, and managed deployment depth to map to commercial plans.

Which features make users upgrade to paid plans in RAG tools?

Users upgrade in RAG tools when they move from proving retrieval quality to operating a real knowledge system. The strongest upgrade triggers are governance, managed deployment, data connectors, evaluation, and scaled document parsing.

Governance creates the clearest upgrade path because it maps directly to organizational risk. Security controls, admin controls, hosted deployment, and enterprise packaging are paid-only in 20 of the 68 tools that include them.

Connectors trigger upgrades when the buyer needs real data rather than demo documents. Since 24 of 44 connector implementations are free limited, the free tier often proves the workflow while paid plans unlock more sources, sync volume, or enterprise integrations.

Evaluation drives upgrades after the system starts affecting real users. RAG quality monitoring, test suites, and observability become more valuable once teams need to compare retrieval changes and catch regressions.

Parsing upgrades are usually tied to volume and document complexity. LlamaParse, Unstructured, Reducto, LandingAI ADE, and similar ingestion-focused products show how document processing can become a paid infrastructure layer rather than a simple free utility.

Agentic workflows can also drive upgrades, but usually through access conditions rather than simple pricing. Their 16 restricted implementations suggest vendors use deployment mode, integrations, or advanced environments as part of the upgrade mechanic.

What should the MVP of a RAG tool include and what should it skip?

The MVP of a general RAG tool should include document ingestion, chunking, embeddings, hybrid retrieval, citations, and basic deployment controls. It should skip knowledge graphs, persistent memory, advanced evaluation, and broad connector depth unless one of those is the workflow anchor.

The non-negotiable MVP surface is the core RAG loop. Parsing, chunking, embeddings, hybrid search, and citations all appear in at least 60 tools, so omitting one creates a visible gap for a general-purpose product.

Basic deployment controls belong in the MVP because the feature is present in 82.9% of the dataset. Even a developer-focused tool needs a credible story for configuration, deployment, or governance, even if the advanced controls come later.

The MVP should include only the workflow anchor that matches the product. A document ingestion tool needs excellent parsing. An evaluation tool needs quality monitoring. A memory and graph retrieval product needs graph or memory depth. A retrieval component needs ranking performance, not a chat UI.

Knowledge graph and memory features should usually be skipped at launch for a broad RAG product. Their penetration is low enough that they differentiate, but not high enough to be required for a first credible version.

Broad connector coverage should also wait unless the product is a managed platform or knowledge chat application. Connectors are table stakes in those workflows, but they add operational complexity that can distract from proving retrieval quality.

What are other interesting feature patterns in RAG tools?

Beyond the headline patterns, RAG tools have several quieter feature dynamics that reveal how the category bundles, hides, and monetizes capabilities.

Citations have one of the largest marketing-versus-packaging gaps in RAG tools. The feature appears in 65 tools, but 26 of those implementations are unclear, which means vendors often claim grounding without making plan-level access obvious.

Reranking has a similar clarity problem. It appears in 50 tools, but 14 of those implementations are unclear, and the ambiguity is especially visible in knowledge chat, memory and graph retrieval, and document ingestion workflows.

Document ingestion tools are more application-adjacent than their category label implies. They have 20 of 23 citation-enabled cases and 17 of 23 governance-enabled cases, even though embeddings and hybrid search are much less common in that workflow.

Memory and graph retrieval tools are not uniformly memory-first. All 8 include knowledge graph features, embeddings, hybrid search, and citations, but only 6 include memory and personalization, which shows that graph retrieval and memory are related but not identical product promises.

Evaluation and observability tools are narrower than they first appear. They all include RAG evaluation, but none include document parsing or data connectors, which makes them add-on infrastructure rather than end-to-end RAG platforms.