Dashboard

Ingest

Add new content to the knowledge base.

Upload files or paste text — the platform breaks it down, identifies the key concepts and relationships, and makes it ready for search and retrieval.

Files are sent directly to the ingestion pipeline (parse → chunk → entity extraction → graph + vector). Synchronous: results appear inline below when complete. Supported: PDF, DOCX, PPTX, XLSX, CSV, HTML, Markdown, AsciiDoc, WebVTT, images, audio, plain text, and source code (.py / .ts / .tsx — pair with the code_structure ontology to build a class/call graph).

Or upload an entire code folder (recurses into subdirectories; filtered to .py, .ts, .tsx for the code_structure ontology):

Knowledge Pipelines

Open Airflow

Airflow runs on Cloud Composer behind Google sign-in, which blocks embedding it inline. Click below to open it in a new tab — you’ll authenticate once with the same Google account that has access to this GCP project.

Open Airflow

Bulk ingest from folder

Bulk import from a folder.

Point to a folder and process every supported file in one go — progress streams live below.

Source

Type a path or click Browse… to navigate the mounted folder tree. The browser is sandboxed under the configured INGEST_BROWSE_ROOT (default /data/ingest).

Leave empty to use the default text-y formats (.txt .md .markdown .rst .yaml .yml .json .html .htm .csv .tsv .log .cypher .cql).

Live event log

    Ontology discovery

    Discover a starting ontology from sample documents.

    Upload a few representative documents and the platform proposes a starter set of concepts and relationships you can review and save.

    PDF, Word, HTML, Markdown, or plain text. Max 10 files, 25 MB each.

    Jobs

    Loading…

    Document library

    All ingested documents, sorted by most recent.

    Drill into any document to see its concepts, relationships, and processing status.

    Loading…

    Quality reports

    Spot quality issues across your knowledge base.

    Surfaces thin concepts, orphans, hubs without depth, lonely nodes, and documents where nothing was extracted.

    Graph connectedness

    How fragmented is the entity graph? Updated daily by a background job — this view never re-runs the algorithm.

    —

    Loading…

    —

    Loading…

    Ontology scorecard

    Health scorecard for each ontology.

    See how much of every defined concept and relationship has real data behind it, plus connectivity and diversity scores.

    —

    Loading…

    Recommendations

    AI-assisted suggestions for improving your ontologies.

    Generation runs in the background — you can leave this tab and come back. Every run is saved to the history below so you can revisit, compare, or print past recommendations.

    History

    Loading…

    Sources

    Connect external data sources.

    Register integrations to systems where your content lives, store credentials securely, and run them on-demand or on a schedule.

    Name Connector Schedule Status Last run Last Run Stats Actions

    Similarity search

    Find content that matches a query by meaning.

    Each result includes the matched text and the concepts associated with it.

    Leave empty to search across all documents. When provided, every key/value must match (AND).

    Context search ⓘLayer Search helps you find the right answer.
    Context Search equips you with the full picture.

    Ask a question and get a grounded answer.

    The platform retrieves relevant content, expands it with related concepts, and answers using only the evidence it found.

    Retrieval runs across the selected layer (or all layers). Without an LLM key, the answer is a deterministic template that surfaces the top retrieved chunk.

    Include extra fields (uncheck for faster response)

    Layer search ⓘLayer Search helps you find the right answer.
    Context Search equips you with the full picture.

    Search across all knowledge layers and merge the results.

    Each fact surfaces once — under the most specific layer that contains it (project > enterprise > domain > general).

    Total candidates fed to the merger = K × number of layers. The merger then dedupes and returns 3–10 bullets.

    Ontology

    Browse the concept and relationship schemas.

    Pick an ontology to see its schema graph plus the full entity and relationship lists.

    Graph

    Run read-only queries against the knowledge graph and explore the result.

    Picking an ontology pre-fills a starter query scoped to it — edit it freely before running.

    Built-ins apply to any ontology. User-saved queries appear here when the matching ontology is selected.

    Read-only — write statements (CREATE / DELETE / MERGE / SET) are rejected by Neo4j. Use parameters by referencing $name; values come from the params dict (advanced — usually empty).

    Type a name and click Save. If the name matches the currently-selected saved query, it’ll be updated; otherwise a new query is created for the selected ontology.

    Graph view

    Nodes coloured by their non-Entity label (Framework / Library / Person / etc). Drag nodes to rearrange. Hover for properties.

    Tabular result

    Import graph

    Load nodes and relationships directly from CSV files. Bypasses the live ingestion pipeline — no chunking, no embeddings, no LLM extraction; only the graph is written.

    Each file's header row decides whether it's a nodes or relationships CSV. Imports are tenant-scoped to the active tenant.

    Upload CSVs

    Drop one or more files. You can mix node and relationship CSVs in a single upload — nodes are imported first so relationships can resolve cross-file references.

    Stamped on rows that don't carry an ontology column.

    Used when a node row has no type column.

    Stamped on every node/edge as a doc_id so a future per-document delete can wipe this import. Auto-generated when blank.

    CSV format reference

    Nodes CSV — header row must include name (or one of: label, title). Recognised columns:

    • node_id / id — stable id used to wire relationships within this upload. Defaults to name.
    • name — required.
    • type / kind / node_type — Neo4j label (e.g. Framework, Person). Defaults to the value above.
    • ontology — defaults to the value above.
    • description — accumulated into n.descriptions[] and surfaced as n.description.
    • Any additional columns are written as node properties.

    Relationships CSV — must include both a source and target id column. Recognised columns:

    • source_node_id / source_id / from — required.
    • target_node_id / target_id / to — required.
    • type / rel_type / relationship — required, e.g. USES, WORKS_AT.
    • ontology — defaults to the value above.
    • Any additional columns are written as relationship properties.

    Identifiers (node labels and relationship types) must match [A-Za-z][A-Za-z0-9_]*. Spaces and dashes are converted to underscores; rows with unsafe labels are skipped.

    Statistics

    Operational metrics across the platform.

    Refresh to recompute.

    Reset data

    Permanently delete all ingested content.

    This cannot be undone. Re-ingesting your sources is the only way to restore content.

    ⚠️ Destructive operation

    Clicking the button below will:

    • Delete the cake-vectors Elasticsearch index (every chunk + embedding).
    • Delete the cake-chunks Elasticsearch index (the keyword index).
    • Run MATCH (n) DETACH DELETE n in Neo4j (every node + every relationship).

    After reset, the indices will be auto-recreated on the next ingest with the correct mapping. Ontologies and layer config are not affected — only your ingested data is wiped.

    Scheduling

    Periodic background jobs.

    Each scheduled run is recorded, and a notification goes to the configured recipient when the job finishes.

    + Create new schedule

    Filled in automatically from the preset; edit only when Custom cron is selected.

    Active schedules

    Loading…

    Token & cost

    Aggregated usage and estimated cost across every workload.

    Costs are estimates derived from per-model rates in the Catalog — edit a rate there to refresh the math.

    LLM Catalog

    Per-model rate table that backs the cost dashboard.

    Costs are computed from input and output token rates; model lookup uses longest-prefix matching against the live model id.

    + Add new model

    Lowercased on save. Longest match wins.

    Catalog

    Tenants

    Each tenant is a customer or workspace.

    Content is scoped to one tenant; ontologies, schedules, and rate catalogs are shared. The active tenant in the sidebar is sent on every request.

    + Create new tenant

    8 chars from [A-HJ-KM-NP-TV-Z2-9] — no 0,1,I,L,O,U. Leave blank to let the system pick.

    All tenants

    Security

    Manage users and the tenants they can access.

    Each user is identified by email. Inactive users keep their audit trail but cannot sign in.

    + Create new user

    All users

    Settings

    Runtime configuration for the platform.

    Edit values inline — most changes take effect immediately.

    Connector catalog (source · data_connectors)

    Available source connector types for ingestion (same JSON as gateway GET /v1/catalog/connectors). Requires AIRFLOW_GATEWAY_BASE_URL.

    Store catalog (destination · stores)

    Available destination store types for artifacts (same JSON as gateway GET /v1/catalog/stores). Requires AIRFLOW_GATEWAY_BASE_URL.

    Loading…

    Labels

    Customise the words the UI uses for shared concepts.

    Each row maps a key (used by the UI markup as data-label="…") to a display name. Edits are saved immediately but only take effect after the next API restart — the running process serves labels from a startup cache.

    LLM Settings

    Configure which LLMs the platform may use and which model handles each operation.

    Section 1 toggles the model allowlist on cake.llm_catalog. Section 2 binds each LLM call site to one of the allowed models. Bindings take effect on the next request.

    1. Allowed models

    Only models ticked here appear in the per-operation dropdowns below.

    2. Per-operation model

    Pick the model each operation should use. Operations with no binding fall back to the global default.

    Prompts

    Library of named prompt templates.

    Each prompt has a unique name code looks up by, plus a category for grouping (search, ingestion, chat, system, or anything else you need).

    + Add new prompt

    Used as a programmatic key. Must be unique.

    Audit logs

    Append-only record of meaningful user and system actions.

    Read-only. Each row carries the actor, tenant, action category, status, request id, and a structured details payload.

    Evals

    Upload a test set and benchmark search quality.

    Each case runs through both search backends and is scored automatically against expected answers — results are captured side-by-side so you can compare head-to-head.

    Runs execute under the active sidebar tenant + the logged-in user. Concurrency is bounded by EVAL_CONCURRENCY (default 3) cases in flight; within a case the two searches fire in parallel. Each search half opens its own Langfuse trace if Langfuse is enabled.

    Start a new run

    Running as … · tenant … · backends layer + context

    Top-level shape: { "metadata": {...}, "questions": [{ id, question, expected_answer, required_facts: [...], category?, difficulty? }] }

    Past runs

    Deep search

    Iterative, agent-driven research across your content.

    Plans sub-questions, runs them in parallel, refines on what it finds, and stitches a synthesised report with citations.

    🔭

    Coming soon

    Retrieval config

    Tune how the platform picks data sources for each kind of question.

    Set per-intent limits and choose which sources are eligible per tenant — no redeploy required.

    ⚙️

    Coming soon

    Data Retrieval

    Customise CAKE to pull from any data source — each one plugs in as a retriever in the catalog below.

    CAKE's intelligent retrieval engine understands the question, decides which sources to query, runs them in parallel, and assembles a single grounded answer. Ideal for agentic platforms: instead of wiring every agent to every backend, point them at CAKE and let it gather what they need.


    Past questions

    Retrievers

    Catalog of data sources the platform can choose from.

    One row per source — built-in, external tool, or custom adapter. Edits take effect on the next request.

    Phase 0: read-only. Add new retrievers via direct SQL insert against cake.retriever_registry for now; admin CRUD lands in Phase 1.

    MCP Servers

    External services the platform can dispatch tool calls to.

    Configure them using the standard JSON format every compatible client uses. Both local and remote servers are supported.

    Graph queries

    Saved query templates that back the Graph panel’s dropdown.

    Built-in templates apply across every ontology; user-saved queries are stored against one ontology. Built-ins can be edited but not deleted.

    + Add new query

    Use {ontology} as a placeholder if you want the query to substitute the currently-selected ontology at fill time (built-in pattern). Leave it out for verbatim queries.

    All saved queries

    API Docs

    Interactive Swagger UI for the FastAPI backend. Use the “Try it out” controls inside each endpoint to call the API live.

    Open in new tab ↗

    MCP (Model Context Protocol)

    Expose retrieval as native tools to compatible clients.

    Auth is via the same API keys you mint from the sidebar.

    Loading…