Admin Console Tools

The Admin Console provides direct access to controlling and monitoring the vector store. Functions are organized in tabs.

1. Search & Write

This tool allows the administrator to manually manage namespaces:

• Search: Search specific namespaces with free text. Filters such as project_id, tags, or metadata help narrow down the results.
• Write: Create new entries directly in a target namespace. Here, ttl_seconds, tags, and metadata JSON can be explicitly set.
• Edit: Existing entries can be corrected or deleted.

Search results are displayed as a separate table below the form.

2. Namespaces

The Namespaces tab shows all existing memory namespaces as a tree, grouped by the segments of the path:

Column	Description
Namespace	The segment this level adds — the full path is in the tooltip
Documents	Number of active entries; on intermediate levels the sum of everything below
Share	Fraction of all documents
Last Modified	Time of the last write operation
Size	Total size of stored content

A filter field above the tree narrows it to matching paths; for vector.user.<uuid> and vector.agent.<uuid> the plain-text name is shown as well and is searchable too.

Clicking a namespace copies it directly into the search field of the Search & Write tab and clears the other form fields there, so the search takes effect immediately.

3. Ranking

The Ranking tab defines how the system handles content from specific namespaces:

• Ranking Bonuses: Increase the relevance of namespaces (e.g., vector.global.knowledge.* receives a bonus of 0.2) so that official knowledge appears above fleeting notes in the search results.
• LLM Instruction Template (Optional): Add specific instructions that are provided to the system prompt when hits from this namespace are found (e.g., “Always cite the source as a link for information from this namespace”).

4. Ingest

Documents are loaded into the vector store via a two-stage process:

Source file (PDF, DOCX, …)
    ↓  Converter
  .md file (reviewable, correctable, LLM-cleanable)
    ↓  Directory Import (Bulk Ingest)
  vector.* Namespace

PDF → Markdown

Converts all .pdf files in a directory into matching .md files in the same directory. The memory store is not written during this step.

Parameters:

• Directory Path – relative to the host process (e.g. ./sources/vector/global/docs)
• OCR Endpoint – optional Apache Tika endpoint (http://host:9998/tika) for image-based PDFs without a text layer
• If .md exists – Replace overwrites existing files, Skip leaves them unchanged

Process:

1. All .pdf files are collected recursively
2. Page-by-page text extraction (pdfjs-dist) with position-aware Markdown reconstruction
3. Headings are detected by font size, tables are rendered as GFM tables
4. Image-only PDFs without a text layer are optionally forwarded to the OCR endpoint
5. The result is written as a .md file next to the source file

After conversion: The .md files can be reviewed, manually corrected, or post-processed by an LLM. They are then loaded into the desired namespace via Directory Import.

Directory Import (Bulk Ingest)

Reads all .md and .txt files from a directory (recursively) and writes them as embeddings into a namespace. Subdirectories are automatically appended as namespace suffixes:

sources/vector/global/docs/       → vector.global.docs
sources/vector/global/docs/api/   → vector.global.docs.api

Parameters:

Parameter	Default	Description
Directory Path	–	Relative to the host process, e.g. ./sources/vector/global/docs
Namespace	–	Target namespace, e.g. vector.global.docs
Chunk Size (Tokens)	512	Maximum size of a chunk in tokens (1 token ≈ 0.75 words)
Overlap (%)	10	Portion of the previous chunk carried over into the next
Chunk Mode	Sliding Window	Sliding Window splits line-by-line with overlap (robust, recommended). Semantic – experimental splits only at Markdown headings (#, ##, …) — useful for well-structured Markdown, but depends on reliable heading detection.
Filter table of contents lines	off	When enabled, lines that look like TOC entries (dotted leaders + page number, e.g. 8.1 Section . . . 64) are removed before chunking. Useful for PDF-converted documents.
If already in memory	Replace	Replace deletes existing chunks for this file in the namespace before writing; Skip skips the file if chunks with this filename already exist

Chunking: Long files are automatically split into overlapping sections. The overlap ensures that context is not lost at chunk boundaries. Each chunk receives metadata (file_name, relative_path, chunk_index, total_chunks, ingested_at).

In Semantic mode, PDF layout noise (page separators ---, HTML comments <!-- page N -->) is stripped before chunking. Each chunk is prefixed with the full heading breadcrumb for LLM context.

Typical workflow after PDF conversion:

1. Convert PDF → Markdown (section above)
2. Review and optionally correct the .md file
3. Run Directory Import → chunks are written to the namespace
4. Optionally use Search & Write to verify the result

5. Audit Log

The audit log records every interaction with the memory:

• Time & Action: When was what done (e.g., read, write, delete).
• Namespace: Which area was affected.
• Detail: Contains technical information about the request, including possible warnings for RLS violations.