Ragent 1.1.0 Ships Hybrid Retrieval and Full-Lineage Tracing for RAG

Hello everyone, I'm Martin.

This 1.1.0 release is the first feature version after 1.0.x. It's not a minor patch, but a systematic upgrade of core modules including the retrieval architecture, document chunking, and prompt management. It's a big step forward from "usable" to "user-friendly, controllable, and traceable."

GitHub: https://github.com/nageoffer/ragent

Let's take a look at the homepage first. It looks like this overall:

Core Highlights

1. Hybrid Retrieval

1.0.x only had vector retrieval. 1.1.0 expands this to a four-channel hybrid: Vector + Elasticsearch keyword + LightRAG graph retrieval + You.com web search. Multiple channels run in parallel and are then fused via RRF and uniformly re-ranked. A Scope layer is introduced to control the knowledge base range, retrieval budgets are configured in three stages, and intent nodes support linking to multiple Collections. New channels are optional to enable; if a remote service goes down, it won't affect local retrieval.

Behind the graph retrieval channel is an automatically constructed knowledge graph:

How Scope defines the retrieval range is clear from this intent recognition tree:

2. Answer Provenance and Recommended Follow-up Questions

Every answer persists source citations, showing exactly which documents and which segments it came from. At the same time, recommended follow-up questions are automatically generated, allowing users to dig deeper along the current topic.

The effect is shown below, with citations and follow-up questions displayed beneath the answer:

3. Document Chunking — Multimodal

Knowledge ingestion forms a Parse → Chunk → Embed → Index pipeline. Parsers are automatically selected by MIME type. Rich documents can be parsed with MinerU. A new VLM image understanding feature performs image-to-text conversion. Vectorized text and display text are decoupled, and each stage has traceable logs.

The knowledge base management in the console, and the corresponding file upload page:

4. Runtime-Configurable Prompts

Moved from editing code and template files to an Agent Profile + Prompt Slot model. Prompts can be maintained and switched directly in the management console, with placeholder validation and fallback to system defaults if unconfigured.

5. Model Routing with Tier Levels

Defines tiers like fast, standard, and deep based on the scenario. Lightweight tasks take the fast tier, while complex answers take the high-capability tier. Configuration validity is checked at startup. Ragent AI's average TTFT (Time To First Token) response is 3-5 seconds.

Combined with full-link tracing, you can see at a glance which nodes a single request passed through and how long each step took:

6. Business Change Auditing

Key management operations record before-and-after snapshots, field differences, the operator, and request context. The management console includes a corresponding query page.

Architecture Level

The retrieval engine, ingestion kernel, prompt runtime, and object storage modules have been refactored into layers. Each module can evolve independently, and adding new retrieval channels or parsers does not intrude on the main flow. Object storage is abstracted behind a unified interface, allowing seamless switching between S3-compatible storage and Alibaba Cloud OSS.

Main Differences from 1.0.x

• Retrieval method: Single-path vector → Four-channel hybrid + RRF fusion
• Answer display: Body text only → Source citations + recommended follow-up questions
• Document Chunking: Opaque process → Four-stage pipeline + MinerU + VLM
• Prompt management: Editing files → Runtime configuration in the management console
• Model routing: Priority list → Tier levels + circuit breaker and degradation
• Observability: Basic logs → Auditing + full-link tracing

Open Source Status

The project is open source on GitHub, and quite a few people have submitted PRs to iterate on it together:

Additionally, Agentic capabilities are being rapidly filled in. The approach uses one framework with two runtime architectures, dynamically switched based on configuration. Refer to the Agentic frontend design page:

Considering that we want everyone to learn more deeply, the relevant code will be released after polishing. Once the Agentic capabilities are released, it will mark Ragent AI entering the 2.0 era.

Final Words

No matter how many AI keywords you put on your resume, an interviewer can see your depth with just a couple of follow-up questions. Ragent lets you talk about AI thoroughly — how hybrid retrieval is fused, how model circuit breakers degrade, how the document chunking pipeline is orchestrated, how evaluation sets quantify effectiveness. Every link is backed by engineering implementation, not just staying at the conceptual level.

The project is continuously iterating. Interested students can clone and run it. If you have any questions, feel free to raise an Issue to discuss.