doc_knowledge_graph.md

Hybrid Knowledge Graph and Dense Retrieval

By integrating a hybrid knowledge graph with dense retrieval, the F1 score in refusal-to-answer task is improved by 2%. For a detailed explanation of the solution, see lark documentation. The essence of this approach is to weight high-frequency words.

This approach is perfectly compatible with the older version. Below are the steps.

1. Build Knowledge Graph

To reduce costs, we use silicon cloud qwen-1.5-110B to extract entity words, and config.ini already supports silicon cloud. The modification is as follows:

[llm.server]
..
remote_type = "siliconcloud"
remote_api_key = "sk-ducerXXXXX"
remote_llm_max_text_length = 40000
remote_llm_model = "alibaba/Qwen1.5-110B-Chat"
rpm = 1000

Assuming the knowledge base is still in the repodir directory, first establish the knowledge graph. After completion, there will be jsonl and pickle files under workdir/kg, and you can test the query.

# About 2 hours
python3 -m huixiangdou.service.kg --build
python3 -m huixiangdou.service.kg --query "How to install mmpose?"
..
+-----------------+-------+------------------------+---------------------------+
|      Query      | State |     Part of Reply      |        References         |
+=================+=======+========================+===========================+
| 如何安装mmpose?  | 0     | repodir/mmpose/READM.. |                           |
|                 |       |                        | <div align="center">      |
|                 |       |                        |   <img                    |
|                 |       |                        | src="resources/mmpose-    |
..

2. Visualization

You can use neo4j for visualization:

python3 -m huixiangdou.service.kg --dump-neo4j --neo4j-uri ${URI} --neo4j-user ${USER} --neo4j-passwd ${PWD}

For more usage, you can use --help

python3 -m huixiangdou.service.kg --help

3. Build Dense Retrieval Feature Library

This step is the feature_store in the README. Since you need to calculate the optimal threshold under hybrid retrieval, do not skip it.

python3 -m huixiangdou.service.feature_store

Test it.

python3 -m huixiangdou.main --standalone

For precision report, see evaluation.