Knowledge base

Knowledge bases are combo agent’s external memory. Load standards, prior projects, and industry references once; the Agent retrieves relevant chunks when answering or generating.

1. Entry point

Use the top Knowledge base tab.

Three areas:

• Top search — by name
• My knowledge bases — created by you
• Team knowledge bases — shared inside the tenant

After switching tenants, the list reflects that tenant’s visibility.

2. Creating a knowledge base

Click Create knowledge base:

Success opens the detail page with tabs: Dataset / Retrieval test / Chunking / Settings.

3. Uploading documents (dataset)

3.1 Supported formats

.pdf   .docx   .xlsx   .txt   .md   .html   .csv   .rtf
.ppt   .pptx
.jpg .jpeg .png .gif .bmp .tiff .webp .svg   (with OCR)
.mp3 .wav   (transcription)
.eml   (email)
.dbc   (automotive CAN DB)

3.2 Upload methods

• Drag-drop files or folders
• Click upload
• Batch: up to 100 files per batch
• Folder recursion keeps paths as tags

3.3 Parse states

4. Chunking (parser_id) — critical

Roughly 70% of “does the KB work” is chunking strategy. Pick parser_id under Settings — 15 options:

Chooser:

4.0.1 Browsing chunks: ABZ ASPICE example

After upload + parse, open a file under Knowledge base → Dataset to inspect chunks.

Using the ABZ KB (Eclipse S-CORE ASPICE bundle) as example.

Step 1: file list

Columns: name, folder, chunk count, date, parser, enable toggle, parse state, actions.

Step 2: open a file

• Breadcrumb KB / Dataset / Chunks
• Each chunk card; per-chunk enable toggle
• Tables stay tabular (e.g. Subject / Program / Platform / Compliance)
• Batch actions: enable / disable / delete
• Preview controls: full text / ellipsis / search / filter

Applies to all parsers (naive, qa, paper, picture, etc.) — granularity differs only.

4.1 Common chunk settings

4.2 Embedding model

Settings → Embedding model:

• bge-large-zh-v1.5 (Chinese default, 1024-dim)
• bge-m3 (multilingual, 1024-dim)
• text-embedding-3-small / -3-large (OpenAI)
• Self-hosted: GPUStack / Ollama / Xinference

5. Retrieval test

Retrieval test tab — validate slices before Agents use them.

• Enter query
• Pick Vector / Text / Hybrid
• Tune top_k (1–30)
• Inspect scores + source doc + chunk IDs

Heuristics:

• Target snippet in Top-5: chunk config OK
• Not in Top-20: shrink chunk size, enable auto questions, or change embedding
• Nothing found: parsing state failed or vectors missing

6. Binding to Agents (essential)

Knowledge bases do not attach to Agents automatically — bind explicitly.

6.1 Three binding modes

Agent template UI may not expose kb_ids yet — use API / cron. A graphical Agent editor checkbox is planned.

6.2 Ephemeral vs formal KB

7. Knowledge graph (parser_id=knowledge_graph)

Adds entity–relation graph:

• Force-directed visualization
• Graph QA (“how are A and B related?”)
• Attribute filters (“entities where type = chip”)

See Skills / ecosystem for related capabilities.

8. FAQ

Q: Uploaded but answers are weak?
A: Check: (1) parsed ✅ (2) retrieval test hits top-5 (3) Agent has kb_ids (4) Plan Mode not stuck in Fast-only skips.

Q: Duplicate uploads?
A: Dedup by hash + filename. Same name overwrites slices.

Q: Disable noisy chunks?
A: Per-chunk enable toggle under Chunks tab.

Q: KB size limits?
A: Soft guidance < ~1M chunks/KB for UI comfort; ES/Infinity can scale further.

Q: Teammate deleted a KB — can I keep using it?
A: No — hard delete. Remove stale kb_ids or replace.

9. Next steps

• Tune memory: Agent settings
• Automotive / patent workflows: respective solution guides
• Sharing: Collaboration