Pipeline, tools, and prompts

The preset Pipeline tab is the runtime assembly surface for a preset. It defines which tools run before retrieval, which retrieval profile is the default, which tools post-process retrieved chunks, which prompts shape the answer LLM, and which tools the answer LLM may call directly.

Where to find this in the Admin Console

Admin Console → Configuration → Presets → open a preset → Pipeline tab.

Runtime flow

When a preset is applied, CoreCube composes a resolved pipeline snapshot. Query-time code uses that snapshot instead of re-reading every tool and prompt row while a user waits for an answer.

The default path is:

1. Run active query tools in position order.
2. Run the preset's default retrieval tool if the query stage did not already produce chunks.
3. Run active chunk tools in position order.
4. Compose the answer system prompt from answer-stage prompt fragments plus every active tool's tool prompt.
5. Send the grounded context and composed prompt to the preset's answer LLM.
6. If the preset exposes retrieval or action capabilities, the answer LLM can call them in a bounded tool-call loop.

Tool stages

Tool type	Pipeline location	What it can do	LLM-callable
query	Before retrieval	Rewrite or expand the user query, then request additional retrieval runs.	No
retrieval	Default retrieval or answer capability	Store a full retrieval profile: candidate pools, fusion weights, freshness floor, reranker, final top-K.	Yes, when not the default retrieval
chunk	After retrieval	Reorder, filter, expand, or annotate candidate chunks before they reach the answer LLM.	No
action	Answer capability	Execute a code tool with an OpenAI-style function schema, scoped to the caller.	Yes

Only one retrieval tool is the default retrieval. Additional retrieval tools are exposed as LLM-callable capabilities, so the answer model can ask for a narrower or wider search when the initial evidence is insufficient.

Retrieval tools store reranker configuration in their settings.reranker_model field. They do not use the generic tool modelId; that field is reserved for tools that need an LLM or executor model.

Prompt stages

Prompt fragments are reusable building blocks. A fragment can be attached to a preset stage or used as a tool prompt.

Stage	Typical fragment types	Runtime use
Query	retrieval_semantic_query, retrieval_keyword_query	Supplied to query tools that rewrite the user's question before retrieval.
Chunk	chunk_selection	Supplied to chunk tools that ask an LLM to classify or expand candidate passages.
Answer	answer, answer_citations, answer_attachments	Composed into the final answer-LLM system prompt.
Tool	tool_prompt	Attached to each tool. Active tool prompts are appended to the answer system prompt.
Ingestion	ocr_extraction	Used by chat-API OCR providers while ingesting scanned or image-based documents.

The answer system prompt is composed in this order:

1. answer fragments in answer-stage position order.
2. Tool prompts for the active default retrieval, retrieval capabilities, query tools, chunk tools, and action capabilities.
3. answer_citations fragments.
4. answer_attachments fragments.
5. A non-editable invariant footer that enforces language, untrusted-context handling, refusal rules, and permission-leak prevention.

Editing a shared fragment affects every preset or tool that references it. The Admin Console shows affected presets and tools before you save.

Seeded default presets

Fresh installs bind the organization to default-balanced. The seeded defaults are:

Preset	Default retrieval	Query tools	Chunk tools	Shape
default-fast	retrieval-fast	None	freshness_decay	Small candidate pools, final top-K 4, lowest latency.
default-balanced	retrieval-balanced	expand_multi_query	freshness_decay	Balanced candidate pools, query expansion, final top-K 7.
default-accurate	retrieval-accurate	expand_multi_query	freshness_decay, llm_select_and_expand	Wide candidate pools, LLM chunk selection, final top-K 12.

These defaults are editable from the preset page. Use Reset to defaults when you want to return one of the seeded presets to its shipped shape, or duplicate a preset before experimenting.

Seeded tools

Tool slug	Type	Purpose
retrieval-fast	Retrieval	Lean retrieval profile for low-latency answers.
retrieval-balanced	Retrieval	Mid-tier retrieval profile used by the fresh-install default.
retrieval-accurate	Retrieval	High-recall retrieval profile with deeper candidate pools.
expand_multi_query	Query	Uses an LLM to create semantic and keyword variants, then fuses the resulting streams.
freshness_decay	Chunk	Re-scores chunks toward recently synced content, bounded by the retrieval freshness floor.
llm_select_and_expand	Chunk	Uses an LLM to select relevant anchors and expand them with neighboring chunks.
noop_chunk	Chunk	Pass-through chunk tool, useful when you need an explicit no-op chunk stage.

Answer model and loop bounds

The answer stage has its own model selection. A preset with tools or prompts but no answer model is incomplete until you pick one.

Two preset-level limits protect runtime tool execution. They apply to deterministic query/chunk tool stages and to LLM-emitted retrieval/action capability calls:

Field	Default	Meaning
max_tool_calls_per_request	8	Maximum tool invocations consumed by one stage budget.
max_tool_call_wall_ms	60000	Maximum wall-clock time spent in one stage budget.

If either bound is exceeded, the request fails fast with a typed tool-loop error instead of running unbounded work. The deterministic query/chunk pipeline and the answer LLM's capability loop each receive their own bounded budget.

Operational notes

• Streaming chat currently supports retrieval and chunk processing before answer generation. If a preset exposes LLM-callable retrieval or action capabilities, use non-streaming chat for that preset until streaming tool-call deltas are supported.
• Tool execution is scoped to the caller's resolved compartments and sensitivity ceiling. A tool cannot widen the API key or user's permissions.
• Code tools execute through the executor sidecar with timeout, memory, and egress allowlist controls.

Related

• System prompts — fragment types, composition order, and examples.
• Retrieval — fields stored inside retrieval tools.
• How retrieval works — the end-to-end query flow.