Skip to content

Knowledge Base ​

The Knowledge Base module is responsible for vector search, Graphify knowledge graph construction, PageIndex indexing, hybrid search and other knowledge management and AI question answering capabilities.


v3.1.7 (Latest) ​

Release Date: 2026-08-28 | Base Version: v3.1.6

1. Overview ​

ItemContent
Release Date2026-08-28
Versionv3.1.7
Base Versionv3.1.6

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Personal Wiki1. Improved the import, organization, query and Q&A flow for personal Wiki materials
2. Users can quickly locate related content around Wiki materials and continue asking questions
Project Knowledge Base1. Added material query support for two project knowledge bases: PPB642 Lovely, PPB850 synergy
2. Supplemented query materials for two project knowledge bases: PPB745 SOATool, PPB800 Syzygium
Cloud Retrieval1. Support querying uploaded project materials from the cloud knowledge base
2. Query results retain the original material source, facilitating users to open the original text and continue confirmation

2.2 Changes ​

ModuleDetails
Material Query1. Unified the usage of personal Wiki, project knowledge base and cloud knowledge base
2. Optimized query result organization, facilitating users to directly view related materials and sources
Release Materials1. Unified the archiving method of project materials, question sets, query results and reports
2. Supplemented this week's project knowledge base usage verification materials, facilitating continuous use in subsequent versions

2.3 Removed ​

ModuleDetails
NoneNo formal features removed in this version

2.4 Bug Fixes ​

ModuleDetails
Wiki Usage Flow1. Fixed the issue of incomplete connection between some Wiki material import and query flows
2. Improved usage consistency between different Wiki material entries
Material Synchronization1. Fixed the issue of inconsistent archiving paths for some project materials
2. Improved source retention for local and cloud query results, facilitating users to trace back to the original text

v3.1.6 ​

Release Date: 2026-08-21 | Base Version: v3.1.5

1. Overview ​

ItemContent
Release Date2026-08-21
Versionv3.1.6
Base Versionv3.1.5

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Knowledge Page Query1. Added an independent query entry for knowledge page materials; users can directly locate candidate articles around LLM Wiki, Karpathy Wiki or Agent Wiki materials and continue Q&A
2. Support candidate retrieval based on wiki document original text, facilitating quick viewing of related knowledge page content
Project Knowledge Base Evaluation1. Added 100-question recall evaluation materials for seven project knowledge bases: PPB685 Graceful, PPB687 academic, PPB715 Energy, PPB718 sandalwood_3M0A, PPB727 decent2024, PPB749 DreamCarStudio-DigitalCar, PPB802 Ocean
2. Evaluation results are uniformly synchronized to kgcase_report, facilitating viewing question sets, hit results, run records and reports by project
Recall Recovery & Synchronization1. Added stability hardening for the recall evaluation recovery pipeline, reducing manual sorting cost when rerunning after interruption
2. Strengthened the evaluation result copying and release process, ensuring question sets, reports and run records remain consistent

2.2 Changes ​

ModuleDetails
Build Stability1. Hardened Chroma persistence validation, covering index files, element count and reopen query checks
2. Optimized exception handling after index build, reducing the risk of incomplete index files being misjudged as successful
Retrieval Result Output1. Kept original hit context and final answer generation separated in local retrieval and distribution flows, facilitating users to view original evidence
2. Optimized source aggregation and result output structure, reducing duplicate information in the retrieval pipeline
Operation Instructions1. Supplemented the knowledge page independent query entry and related instructions, unified the usage of different knowledge base paths
2. Updated recall evaluation and release process instructions, facilitating subsequent versions to synchronize materials along unified paths

2.3 Removed ​

ModuleDetails
NoneNo formal features removed in this version

2.4 Bug Fixes ​

ModuleDetails
Evaluation Synchronization1. Fixed unstable source file handling during recall evaluation result copying and recovery, ensuring reports, question sets and run records are synchronized completely
2. Fixed issue of inconsistent paths of some evaluation artifacts requiring manual completion during release archiving
Index Persistence1. Fixed Chroma persistence validation boundary conditions, improving queryability of reopened indexes under Windows environment
2. Fixed issue of insufficient validation information in some build flows, facilitating locating index persistence exceptions

v3.1.5 ​

Release Date: 2026-08-14 | Base Version: v3.1.4

1. Overview ​

ItemContent
Release Date2026-08-14
Versionv3.1.5
Base Versionv3.1.4

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Project Knowledge Base Evaluation1. Added 100-question recall evaluation materials for seven project knowledge bases: PPB682 decent, PPB687 academic, PPB715 Energy, PPB727 decent2024, PPB737 Citrus, PPB783 soul, PPB836 Realization
2. Evaluation results are uniformly synchronized to kgcase_report, facilitating users to view question sets, hit results, run records and reports by project
RAG Workbench Reports1. Added RAG workbench evaluation reports and HTML details for two knowledge bases: herdsman and starry2
2. Both knowledge bases completed 100-question verification, with document hit Recall@5 both at 100.00%
Knowledge Graph Prompts1. Support specifying custom prompts during knowledge graph construction, facilitating different projects to adjust graph extraction focus by material type
2. Retain the default graph construction entry; when not specified by users, the standard behavior continues to be used

2.2 Changes ​

ModuleDetails
Recall Evaluation Process1. Optimized the recall evaluation script, supporting synchronizing project evaluation results to the unified report directory
2. Improved question set optimization, evaluation record copying and report rendering flows, reducing manual sorting path differences
Build Stability1. Hardened Chroma persistence validation, covering HNSW files, index element count and reopen query checks
2. Optimized validation failure handling after index build, reducing the risk of incomplete index files being misjudged as successful
Retrieval Result Output1. Optimized result deduplication, source aggregation and output structure in local retrieval and distribution flows
2. Kept retrieval context and final answer generation separate, facilitating users to view original hit evidence

2.3 Removed ​

ModuleDetails
NoneNo formal features removed in this version

2.4 Bug Fixes ​

ModuleDetails
Evaluation Data Synchronization1. Fixed unstable source file handling issue during recall evaluation result copying, ensuring reports, question sets and run records are synchronized completely
2. Fixed issue of inconsistent paths of some evaluation artifacts requiring manual completion during release archiving
Index Persistence1. Fixed Chroma persistence validation boundary conditions, improving queryability of reopened indexes under Windows environment
2. Fixed issue of insufficient validation information in some build flows, facilitating locating index persistence exceptions

v3.1.4 ​

Release Date: 2026-08-06 | Base Version: v3.1.3

1. Overview ​

ItemContent
Release Date2026-08-06
Versionv3.1.4
Base Versionv3.1.3

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Knowledge Graph Capability1. Added explicit knowledge graph construction and query entry; when users explicitly need a graph, they can build a graph independently and query by graph, vector or hybrid mode
2. Added a graph directory in knowledge base artifacts, facilitating entity-relationship queries in addition to original document retrieval
Project Knowledge Base Evaluation1. Added 100-question recall evaluation materials for two project knowledge bases: PPB737 Citrus and PPB783 soul
2. Both projects retain question sets, retrieval results, execution records and reports, facilitating subsequent review and horizontal comparison
Upload & Release Process1. Supplemented instructions for uploading a knowledge base by current working directory or specified project
2. Supplemented overwrite update, post-upload activation and knowledge base marketplace package generation flows, facilitating knowledge base delivery and reuse

2.2 Changes ​

ModuleDetails
Build Process1. Support batch-based build and resume after interruption, reducing duplicate processing cost after large knowledge base build failure
2. Enhanced index persistence check after build completion, and compatible with knowledge base construction under Windows Chinese paths
Retrieval & QA Process1. Adjusted local retrieval output to first return raw hit context and source information, then Skill organizes the final answer according to unified specification
2. Optimized retrieval result deduplication, source aggregation and cloud knowledge base source link display, facilitating user traceability to original materials
Runtime Configuration & Logs1. Unified runtime configuration reading order, prioritizes SDW global settings, and retains project configuration and environment variable fallback
2. Optimized build log and input file display, reducing interference from long lists and progress output on user result checking

2.3 Removed ​

ModuleDetails
Old Retrieval Entry Cleanup1. Cleaned up old single-file retrieval entry, consolidated retrieval calls to current query command and dispatch flow
2. Removed unnecessary retrieval intermediate output, reducing execution differences caused by script entry mixing

2.4 Bug Fixes ​

ModuleDetails
Build Stability1. Fixed issue of continuing execution after incomplete index file persistence in some environments
2. Fixed issue of possible failure when building vector index under Chinese paths
Retrieval Stability1. Fixed issue of still triggering local model validation when only retrieving cloud knowledge base
2. Fixed several boundary issues in retriever cache, result sorting and source information display

v3.1.3 ​

Release Date: 2026-07-31 | Base Version: v3.1.2

1. Overview ​

ItemContent
Release Date2026-07-31
Versionv3.1.3
Base Versionv3.1.2

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Evaluation Workbench1. Added rag-eval-workbench knowledge base evaluation Skill, supports generating evaluation sets based on source evidence, executing TopK retrieval and outputting HTML/JSON reports
2. Added KMS 100-question evaluation samples, report contracts and LLM judge reference materials, unifying subsequent knowledge base evaluation artifact format
Project Knowledge Base Evaluation1. Added Top5 recall evaluation reports for two project knowledge bases: PPB745 SOATool and PPB748 Elegant
2. Added build and evaluation artifacts for pdc-ii and delight knowledge bases, and supplemented RAG workbench reports for clearly, surmounter and tuberose
Statistics Dashboard1. Added knowledge base work summary HTML, owner configuration and knowledge base capability inventory materials
2. Added this week's capability curve, statistics for 7 knowledge bases by main metrics, current cumulative average is 88.00%

2.2 Changes ​

ModuleDetails
Evaluation Report Format1. Unified rag_eval_report_top5.html and rag_eval_report_top5.json output, supplemented Doc Hit@K, Chunk Recall@5, MRR@5 and Answer Token F1 metrics
2. Adjusted report structure for clearly, surmounter, tuberose, pdc-ii and delight, facilitating horizontal comparison and subsequent trend tracking
Retrieval & Evaluation Pipeline1. Integrated BM25/vector hybrid retrieval evaluation pipeline, and organized metric change summary for three types of knowledge bases
2. Packaged project evaluation CSV to aikokg-build-recall-eval, supplemented source selection and recall evaluation instructions
Build Optimization1. Added PPB745 build duration, token consumption and recall impact quality constraint optimization verification
2. Added metadata inference cache optimization verification, incremental build can reuse cache to reduce duplicate LLM inference

2.3 Removed ​

ModuleDetails
Old Evaluation Artifact Cleanup1. Cleaned up historical directories in milestone knowledge base test report V1 that have been replaced by new statistics entry
2. Removed old render_kb_eval_html.py and expired data format analysis/filter scripts, consolidated to new evaluation workbench and build flow scripts

2.4 Bug Fixes ​

ModuleDetails
Report Metrics1. Fixed issue of missing Top5 metrics, HTML details and JSON results in some evaluation reports
2. Improved knowledge base effect change tracking and detail table display, reducing metric misalignment during manual summary
Build Stability1. Fixed time and token waste issue caused by duplicate inference in metadata incremental build scenario
2. Optimized handling of low-value inputs such as duplicate JSON/Markdown and lockfile, reducing build noise while keeping recall quality controllable

v3.1.2 ​

Release Date: 2026-07-24 | Base Version: v3.1.1

1. Overview ​

ItemContent
Release Date2026-07-24
Versionv3.1.2
Base Versionv3.1.1

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Runtime Configuration1. Added aikokg runtime configuration externalization capability, unified management of cloud addresses, models, build parameters and retrieval parameters
2. Added cloud service address fallback mechanism, supports switching to backup base url when default service is unavailable
Recall Evaluation1. Added Top5 recall reports for four project knowledge bases: PPB681, PPB717, PPB729, PPB800, each includes 100 cases of questions.jsonl, results.json, report.md and run_manifest.json
2. Added unified QA evaluation reports for clearly, surmounter and tuberose knowledge bases, supplemented Recall@1, Recall@5, Token F1 and insufficient evidence statistics
Data Analysis1. Added raw data format analysis scripts and result files for projects such as clearly, surmounter, tuberose, KCarOS-AD, Elegant
2. Added knowledge base build statistics summary, providing data basis for subsequent multi-project build strategies and thread parameter optimization

2.2 Changes ​

ModuleDetails
Build & Retrieval Pipeline1. Refactored knowledge base build, parsing, indexing, retrieval and routing flows, optimized multi-module collaboration
2. Optimized dynamic thread, build statistics, hybrid retrieval and local retrieval logic, reducing multi-project build and retrieval configuration cost
Cloud & Source Links1. Adjusted aikokg default configuration, prioritizes reading SDW runtime settings
2. Supported KMS DingTalk source links, improved availability of retrieving back to original materials
Evaluation Output1. Adjusted evaluation report output format, consolidated multi-knowledge base evaluation reports to kgcase_report unified entry
2. Updated project recall reports and evaluation question sets, kept report, result and test question paths consistent

2.3 Removed ​

ModuleDetails
Historical Demo Cleanup1. Cleaned up skills/hmi_demo historical demo project and expired PageIndex intermediate results
2. Cleaned up old Graphify and vector recall temporary test artifacts under milestone directory, avoiding mixing with current formal evaluation entry

2.4 Bug Fixes ​

ModuleDetails
Build Process1. Fixed print and statistics variable name errors in build.py
2. Fixed several boundary issues in build, parsing, indexing and local retrieval flows after evaluation question optimization
Configuration Stability1. Fixed runtime configuration priority issue, ensuring SDW settings take effect
2. Fixed base url reading and fallback handling in cloud retrieval, upload and routing, reducing environment switching failure risk

v3.1.1 ​

Release Date: 2026-07-17 | Base Version: v3.1.0

1. Overview ​

ItemContent
Release Date2026-07-17
Versionv3.1.1
Base Versionv3.1.0

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Local Knowledge Base Skill1. Added complete local knowledge base build and retrieval skill package, and consolidated aikokg-new into the formal aikokg skill directory
2. Added local knowledge base QA generation capability, completing build, retrieval, registration, upload and release script pipeline
Evaluation & Demo1. Added milestone knowledge base multi-project evaluation question sets, result files and summary reports
2. Added HMI data parsing work report, and attempted to enhance MinerU parsing flow via agent
Planning Materials1. Added Q3/Q4 scheduling, asset assessment and construction metric review materials for project knowledge bases in the past three years
2. Added NAS project knowledge base archive asset list and assessment details

2.2 Changes ​

ModuleDetails
Build Process1. Unified embedding build mode, adjusted build, indexing, parsing, metadata and main entry scripts
2. Optimized parallel parsing, loading thread count and embedding batch configuration, improving local build throughput
Skill Release1. Adjusted skill release default visibility to public
2. Updated operation instructions, supplemented Tsinghua pip source installation command and fixed QA routing rule description
Evaluation Results1. Updated evaluation results for ASPICE, FCSM, KMS, Trollius, Sprout, Ascension, koi and other knowledge bases
2. Adjusted manual recall evaluation report structure, removed redundant statistical fields and organized test report baselines

2.3 Removed ​

ModuleDetails
Historical Directory Cleanup1. Removed old scattered skills/aikokg submodule directories, switched to unified skill directory maintenance
2. Cleaned up expired manual test set JSON and old evaluation report intermediate files

2.4 Bug Fixes ​

ModuleDetails
Configuration & Release1. Fixed subprocess output encoding and null value handling issues in publish_skill.py
2. Added SDW settings support, and updated OpenAI API base address configuration
Parsing Stability1. Added PDF parser warning suppression, reducing parsing log noise
2. Fixed embedding batch default value and path description, reducing build and release configuration misuse

v3.1.0 ​

Release Date: 2026-07-03 | Base Version: v3.0.6

1. Overview ​

ItemContent
Release Date2026-07-03
Versionv3.1.0
Base Versionv3.0.6

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Knowledge Base Upload & Registration1. Added local knowledge base build result upload to cloud service capability
2. Registration process supports automatic SKILL.md generation, and supplements skill reference and operation instructions
Parsing & Demo1. Added Excel file parsing support, and synchronously adjusted default chunk configuration
2. Added HMI PageIndex illustrated QA demo and V2 frontend, QA service, PageIndex build logic

2.2 Changes ​

ModuleDetails
Agentic RAG1. Refactored knowledge base routing and retrieval logic, supports directly specifying knowledge base scope
2. Adjusted HTTP API, MCP, Streamlit entry and registry initialization flow, removed old self-test data dependencies
Retrieval Pipeline1. Optimized local retrieval flow, vector retriever and keyword retrieval collaboration
2. Unified skill script root directory to skills, fixed upload exclusion directory and retrieval configuration
Planning & Documentation1. Added 2026 Q3 knowledge base milestone planning, weekly OKR and prompt formulation summary
2. Updated KMS knowledge base metadata file path, operation instructions and skill reference documentation

2.3 Removed ​

ModuleDetails
Agentic RAG Historical Dependencies1. Removed old kb_registry sample library files
2. Removed historical selftest data and old self-test scripts, consolidated to new registry and interface verification flow

2.4 Bug Fixes ​

ModuleDetails
Path & Default Configuration1. Fixed path handling logic in upload knowledge base interface
2. Fixed OPENAI_LLM_METADATA_MODEL default value description, and updated KMS metadata path references
Retrieval Stability1. Fixed some retrieval configuration and distribution flow issues
2. Optimized local retrieval fallback and knowledge base scope control, reducing routing misselection risk

v3.0.6 ​

Release Date: 2026-06-26 | Base Version: v3.0.4

1. Overview ​

ItemContent
Release Date2026-06-26
Versionv3.0.6
Base Versionv3.0.4

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Unified Evaluation & Samples1. Added kg_case project knowledge base unified evaluation samples and latest fixed result entry
2. Added Trollius PageIndex local evaluation script and 100-question evaluation result output
Skill & Demo1. Added complete local knowledge base skill package, knowledge base agent API service and project entry documentation
2. Added DOCX document RAG application demo based on LlamaIndex

2.2 Changes ​

ModuleDetails
Retrieval Pipeline1. Refactored local retrieval logic, split vector retrieval and keyword retrieval paths
2. Enhanced cloud health detection and fallback mechanism, kept retrieve_dispatch.py as unified distribution entry
Build & Process1. Optimized knowledge base build flow, configuration and project knowledge base processing flow
2. Updated aikokg documentation, Skill constraints and weekly sync flow instructions

2.3 Removed ​

ModuleDetails
Evaluation & Intermediate Artifact Cleanup1. Cleaned up expired evaluation reports and conversion-generated historical intermediate files
2. Cleaned up irrelevant test data, consolidated weekly sync archive entry

2.4 Bug Fixes ​

ModuleDetails
Preprocessing & Parsing1. Optimized incremental update, batch concurrency and parsing flow, fixed stability issues in knowledge base processing pipeline
2. Fixed several exception scenarios in modified document update and preprocessing flow

v3.0.4 ​

Release Date: 2026-06-12 | Base Version: v3.0.3

1. Overview ​

ItemContent
Release Date2026-06-12
Versionv3.0.4
Base Versionv3.0.3

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Incremental Build1. Added document-level incremental build and incremental update workflow
2. Added incremental_state.json state recording, supports new and deleted document identification
Retriever1. Added PageIndex retriever structure and local search capability
2. Enhanced page-level index construction, query examples and industry professional capability research

2.2 Changes ​

ModuleDetails
Registration Process1. Added kb_root and source_root configuration fields to knowledge base registration process
2. Unified build mode and registration process, simplifies self-testing and default behavior
Interface Updates1. Updated cloud search interface address, request path and tool name
2. Updated path configuration, default behavior and standard customer problem responses for aikokg-related documentation

2.3 Removed ​

ModuleDetails
NoneNone

2.4 Bug Fixes ​

ModuleDetails
Path & Configuration1. Fixed inconsistent cloud knowledge base request path and tool name
2. Optimized PageIndex search skill, skill port configuration and document structure

v3.0.3 ​

Release Date: 2026-06-05 | Base Version: v3.0.2

1. Overview ​

ItemContent
Release Date2026-06-05
Versionv3.0.3
Base Versionv3.0.2

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Build Scripts1. Added three types of knowledge base build scripts: PageIndex, Graphify, vector library
2. Added unified parsing and preprocessing capabilities from Office/PDF to Markdown
Search Capability1. Added local/cloud knowledge base routing, distribution and query capabilities
2. Added LLM routing decision, hybrid search and cloud HTTP search invocation capabilities

2.2 Changes ​

ModuleDetails
Metadata Specification1. Unified knowledge base metadata file naming to metadata.json
2. Enhanced register_kb registration process, supplements frontend manifest generation and metadata interface self-test
Documentation Updates1. Updated aikokg skill description, user manual and cloud search rules
2. Fixed packaging paths, naming rules and checklists in weekly sync Skill

2.3 Removed ​

ModuleDetails
NoneNone

2.4 Bug Fixes ​

ModuleDetails
Parsing & Building1. Fixed Markdown outer code block wrapping, process document mis入库 and parsing fallback issues
2. Optimized Graphify CLI automatic installation, PageIndex batch processing and directory filtering logic

v3.0.2 ​

Release Date: 2026-05-29 | Base Version: v3.0.1

1. Overview ​

ItemContent
Release Date2026-05-29
Versionv3.0.2
Base Versionv3.0.1

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Metadata Interface1. Added knowledge base metadata acquisition interface capability
2. Added MCP metadata self-test script and documentation
Routing Evaluation1. Added intelligent hybrid routing hit rate evaluation results
2. Supplemented knowledge base management instructions and maintenance materials

2.2 Changes ​

ModuleDetails
Release Process1. Updated weekly sync Skill release process and checklist
2. Clarified Release Note, test archiving and agentic_rag packaging rules
Patent Documentation1. Completed patent disclosure for knowledge base cloud management and local collaboration solution
2. Completed patent disclosure for adaptive selection construction scheme based on data source characteristics

2.3 Removed ​

ModuleDetails
NoneNone

2.4 Bug Fixes ​

ModuleDetails
Version Alignment1. Regularized v3.0.2 release product path and naming
2. Corrected version alignment relationship between package, archive and Release Note

v3.0.1 ​

Release Date: 2026-05-27 | Base Version: v3.0.0

1. Overview ​

ItemContent
Release Date2026-05-27
Versionv3.0.1
Base Versionv3.0.0

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Test Framework1. Added knowledge base algorithm test report
2. Added five types of solution test indexes: vector-db, Graphify, GraphRAG, HybridQA, PageIndex
3. Added PageIndex high recall and RAGAS low cause analysis
Archive Management1. Generated v3.0.2 test case archive package
2. Generated v3.0.2 test report archive package

2.2 Changes ​

ModuleDetails
Directory Specification1. Unified active research workspace to knowledge-base-research/
2. Unified archive root directory to archive/
3. Archive materials are usually manually migrated every Friday
Rule Synchronization1. Synchronized team AGENTS rules
2. Does not force all results to additionally output Markdown
3. Prioritizes directory specification when generating content

2.3 Removed ​

ModuleDetails
NoneNone

2.4 Bug Fixes ​

ModuleDetails
Path References1. Fixed old path references from docs/... to archive/... in documentation
2. Cleaned up invalid paths in active workspace related_files

SDW AIKO Product Documentation