Backend · GitHub · Interview

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

By 马丁玩编程 · Aug 19, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most RAG frameworks stop at vector similarity. Ragent 1.1.0 gives teams a production-grade retrieval stack that combines four search strategies, traces every answer back to its source, and surfaces the cost of each retrieval step — all configurable at runtime without touching code. For anyone building a customer-facing knowledge bot or internal Q&A system, these are the observability and control primitives that separate a demo from a deployable service.

Summary

Ragent 1.1.0 is a systematic rebuild of the retrieval, ingestion, and prompt-management layers. Its hybrid retrieval now runs vector, Elasticsearch keyword, LightRAG graph, and You.com web search in parallel, merging results through Reciprocal Rank Fusion and a unified re-ranker. A Scope layer constrains which knowledge bases a query hits, and retrieval budgets are configurable across three stages. Every answer carries persistent source citations down to the document segment, plus auto-generated follow-up questions. The document chunking pipeline now handles multimodal content: a VLM converts images to text, MinerU parses rich documents, and the pipeline logs each stage for traceability. Prompts move from template files to a runtime-configurable Agent Profile and Prompt Slot model with placeholder validation and system-default fallbacks. Model routing is organized into tiered profiles (fast, standard, deep) with circuit-breaker degradation, and a full-link trace shows per-node latency for every request. The architecture refactors retrieval, ingestion, prompt runtime, and object storage into independently evolvable modules, with object storage abstracted behind a single interface that swaps between S3-compatible stores and Alibaba Cloud OSS.

Takeaways

— Retrieval is now a four-channel hybrid: vector, Elasticsearch keyword, LightRAG graph, and You.com web search run in parallel and are fused via RRF plus a re-ranker.

— A Scope layer maps user intent to specific knowledge-base Collections, and retrieval budgets are configurable in three stages.

— Every answer includes persistent source citations pointing to the exact document segments used, plus automatically generated follow-up questions.

— The document ingestion pipeline (Parse → Chunk → Embed → Index) selects parsers by MIME type, uses MinerU for rich documents, and applies a VLM to convert images into searchable text.

— Prompts are managed through an Agent Profile and Prompt Slot model in the admin console, with placeholder validation and automatic fallback to system defaults.

— Model routing uses tiered profiles (fast, standard, deep) with circuit-breaker degradation; average time-to-first-token is 3–5 seconds.

— Full-link tracing captures every node and its latency per request, and a business-change audit log records before/after snapshots, field diffs, operator identity, and request context.

— The architecture splits retrieval, ingestion, prompt runtime, and object storage into independent layers, with object storage abstracted behind a unified S3/OSS interface.

— Agentic capabilities are in development using a single framework with two runtime architectures switched by configuration, targeting a 2.0 release.

Conclusions

Shipping a four-channel hybrid retrieval system with RRF fusion and a re-ranker moves Ragent past the typical single-vector demo into territory that usually requires stitching together separate services.

Making the Scope layer configurable per intent node and Collection, rather than a global setting, gives operators fine-grained control over which knowledge base answers which question — a practical necessity for multi-tenant or multi-domain deployments that most RAG frameworks skip.

Decoupling vectorized text from display text in the chunking pipeline is a quiet but important design choice: it lets the embedding model consume one representation while the user sees another, which matters when images or rich formatting are involved.

The tiered model-routing with circuit-breaker degradation treats LLM calls as a reliability problem, not just a cost problem, which is the right posture for any system that must stay online when a model endpoint goes down.

Concepts & terms

RRF (Reciprocal Rank Fusion)

A method for combining ranked result lists from multiple retrieval systems. It scores each document by the reciprocal of its rank in each list and sums those scores, producing a single fused ranking without requiring relevance scores to be calibrated across systems.

LightRAG

A retrieval-augmented generation approach that builds a knowledge graph from documents and retrieves information by traversing graph relationships, rather than relying solely on vector similarity. It captures entity-level connections that vector search can miss.

VLM (Vision-Language Model)

A model that takes images as input and produces text descriptions. In Ragent's chunking pipeline, a VLM converts images embedded in documents into searchable text so they can be indexed alongside the document's written content.

TTFT (Time To First Token)

The latency from when a request is sent to an LLM until the first token of the response is generated. It is a key metric for perceived responsiveness in conversational AI systems.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗