Open-Source Agent vs. 1C Napar: Hybrid RAG on DeepSeek Wins in Practice

The agent is built on LangGraph (Deep Agents) with full-text search via PostgreSQL FTS and vector search via pgvector + HNSW. Results are merged using Reciprocal Rank Fusion. A metadata parser navigates 236,000 files and 193,000 directories across standard 1C configurations. For web search, the team uses the Yandex API instead of chat-builtin search engines.

• Managed lock timeouts: the agent proposed a filter for TLOCK events (>1 sec), TTIMEOUT, and TDEADLOCK, plus an analysis of the blocking connection's actions; Napar offered a basic configuration without details
• Connecting a reference catalog to BSP: the agent pointed to defined types, the WriteObjectVersion subscription, and a manager handler; Napar suggested an alternative via the platform's built-in data history
• Blocking posting cancellation: the agent provided ready-made code with a query to an accumulation register and warned about the need for a managed lock

The obvious conclusion: a vendor has full access to its materials and unlimited context, so its model must always be better. But that's not the case. In practice, the quality gap between a specialized model and a regular cloud model often doesn't justify the cost of fine-tuning — provided the model is backed by good indexing and the right knowledge sources.

Habr took real questions from 1C developers and compared an open-source agent against the commercial Napar assistant. The results were unexpected.

Architecture: Hybrid RAG on LangGraph

The models chosen were top-tier: deepseek-v4-flash, gpt-6.1-sol, glm-5.3-flash. All handle tools equally well and have sufficiently large context windows. It turned out the model wasn't the critical factor — its environment was.

The framework is LangGraph (Deep Agents), with planning, subagent delegation, and context control via history summarization.

For knowledge retrieval, the team implemented a hybrid RAG. The full-text branch uses PostgreSQL FTS with a Russian dictionary; the vector branch uses pgvector with an HNSW index. They're merged via Reciprocal Rank Fusion, which boosts high-ranking candidates from each result set to the top of the final list.

For standard configuration source code, they built a fast index covering 236,000 files and 193,000 directories. They also added a metadata parser that collects objects, attributes, and the relationships between them — helping the agent navigate modules.

For internet search, they use the Yandex API instead of chat-builtin search. This yields more relevant results for Russia-specific content. The agent aggregates and summarizes everything it finds.

Three Tests on Real-World Tasks

1. A technology log for managed lock timeouts.

The agent produced a configuration with a filter for long waits (TLOCK events over 1 second, TTIMEOUT, TDEADLOCK), analyzed the WaitConnections field to identify the blocking connection, and suggested tracking its actions: long-running queries, calls to external services, prolonged processing within a transaction.

Napar also provided a configuration and mentioned WaitConnections, but focused on gathering materials for an investigation rather than analyzing the blocking side.

2. Connecting a non-standard reference catalog to the BSP versioning subsystem.

The agent: add the type to the defined types VersionedData and VersionedDataObject, implement a WhenDefiningObjectVersioningSettings handler in the catalog manager, and call initialization from the form's OnCreateAtServer event. The WriteObjectVersion subscription already listens for the needed event. It provided code and specified that catalogs use this exact subscription, not WriteDocumentVersion.

Napar proposed a different mechanism: the catalog's Data History property and a scheduled job calling DataHistory.UpdateHistory(). Valid, but it's an alternative — not a connection to the standard BSP subsystem.

3. Blocking posting cancellation when batches have already been written off.

The agent provided a BeforeUndoPosting handler with a query against the GoodsByBatches register, a filter on outgoing movements, and a check via QueryResult.Empty(). It also included an important warning: preventing simultaneous write-offs requires a managed lock (absent from the example).

Napar provided a schema without going into the details of the different scenarios.

Conclusion

An open model paired with quality indexing — hybrid RAG, metadata parsing, and the right choice of sources — often performs no worse than a narrowly specialized model. It's more cost-effective to invest in search infrastructure than in fine-tuning. 1C Napar handles standard questions fine but gives shallower answers on complex ones.