Tool Calling

A Tool Calling is a function that the LLM can invoke autonomously during a conversation to retrieve information or perform an action. Instead of relying purely on training data, the LLM calls tools to fetch live data, search indexed content, execute code, and more, then incorporates the results into its reasoning.

Turing ES includes native tools organized into categories, plus support for external tools via MCP Servers. Tools are enabled per AI Agent, each agent selects only the tools it needs.

---

DSL Search: 6 tools

These tools provide Elasticsearch-compatible Query DSL access to any Semantic Navigation Site. Inspired by the Elasticsearch MCP Server, they allow the LLM to build and execute full DSL queries, aggregations, and analytics, automatically translated to the site's configured search engine (Elasticsearch, Solr, or Lucene).

Tool	Description
dsl_list_indices	Lists all available SN Sites with engine type, locales, and status
dsl_get_mappings	Returns field mappings (types, facets, multi-valued) in Elasticsearch format
dsl_search	Executes a full Elasticsearch Query DSL search (40 query types, 35 aggregations)
dsl_get_document	Retrieves a document by ID with all stored fields
dsl_suggest	Autocomplete suggestions and spell-check corrections
dsl_get_shards	Shard/core information: document counts, engine type, endpoints, locales

How the LLM uses DSL tools

The LLM follows a pattern similar to how a developer uses the Elasticsearch API:

1. Discover → dsl_list_indices("*") to find available sites
2. Understand schema → dsl_get_mappings("mySite") to see fields and types
3. Search → dsl_search("mySite", "en", queryBody) with full DSL
4. Deep dive → dsl_get_document("mySite", "en", "doc-123") for full content
5. Correct typos → dsl_suggest("mySite", "en", "enterprse") for corrections

The dsl_search tool

This is the most powerful tool. The queryBody parameter accepts the complete Elasticsearch _search request body as a JSON string. The LLM constructs the query based on the user's intent:

Simple search:

{"query":{"match":{"title":"machine learning"}},"size":5}

Filtered search with aggregations:

{
  "query":{"bool":{"must":[{"match":{"title":"security"}}],"filter":[{"term":{"type":"article"}}]}},
  "size":10,
  "aggs":{"by_author":{"terms":{"field":"author","size":5}}},
  "highlight":{"fields":{"title":{},"body":{}}}
}

Date range with sorting:

{
  "query":{"range":{"date":{"gte":"2025-01-01"}}},
  "sort":[{"date":"desc"}],
  "size":20
}

Similar documents (More Like This):

{"query":{"more_like_this":{"fields":["title","body"],"like":"enterprise search platform","min_term_freq":1}}}

Facet value discovery (terms aggregation):

{"query":{"match_all":{}},"size":0,"aggs":{"categories":{"terms":{"field":"category","size":50}}}}

For the complete DSL reference, see DSL Query API and DSL Compatibility Matrix.

Prompt examples

• "Search for documents about authentication in the Sample site" → dsl_search with match query
• "What fields can I filter by on the Sample site?" → dsl_get_mappings
• "Show me the full content of document ID abc-123" → dsl_get_document
• "How many documents do we have per content type?" → dsl_search with terms aggregation
• "Find documents similar to the article about Solr configuration" → dsl_search with more_like_this
• "What articles were updated in the last 7 days about security?" → dsl_search with range + sort
• "Compare the features of Product A and Product B" → dsl_get_document called for each product
• "What categories are available?" → dsl_search with size:0 and terms aggregation
• "How many documents per locale does the site have?" → dsl_get_shards

---

RAG / Knowledge Base: 4 tools

These tools enable the LLM to query the Knowledge Base, files indexed from MinIO via the Assets section, using semantic similarity search.

Tool	Description
search_knowledge_base	Searches for relevant documents by semantic similarity (top 10, threshold 0.7)
knowledge_base_stats	Returns statistics: total files, chunks, and storage size
list_knowledge_base_files	Lists all indexed files, with optional keyword filter
get_file_from_knowledge_base	Retrieves the full indexed content of a specific file

Prompt examples:

• "What does our internal documentation say about deployment procedures?" → triggers search_knowledge_base
• "How many files are indexed in the knowledge base?" → triggers knowledge_base_stats
• "Show me the full content of the file onboarding-guide.pdf" → triggers list_knowledge_base_files then get_file_from_knowledge_base

---

Web Crawler: 2 tools

Tool	Description
fetch_webpage	Fetches a web page by URL and returns its content as plain text
extract_links	Extracts all links from a web page, with optional keyword filter

Prompt examples:

• "What does the Apache Solr documentation say about faceted search?" → triggers fetch_webpage
• "List all links on our company blog homepage" → triggers extract_links

---

Finance: 2 tools

Tool	Description
get_stock_quote	Current price and market data for a stock ticker symbol
search_ticker	Looks up a ticker symbol by company name or keyword

Prompt examples:

• "What is the current stock price of Apple?" → triggers search_ticker then get_stock_quote
• "Show me the market data for MSFT" → triggers get_stock_quote

---

Weather: 1 tool

Tool	Description
get_weather	Current weather and forecast for a city (1–7 day range)

Prompt examples:

• "What's the weather forecast for São Paulo this week?" → triggers get_weather

---

Image Search: 1 tool

Tool	Description
search_images	Searches the web for images and returns URLs and descriptions

Prompt examples:

• "Find images of enterprise search architecture diagrams" → triggers search_images

---

DateTime: 1 tool

Tool	Description
get_current_time	Returns the current date and time for a given IANA timezone

Prompt examples:

• "What time is it right now in Tokyo?" → triggers get_current_time

---

Code Interpreter: 1 tool

Tool	Description
execute_python	Executes Python code in a sandboxed environment and returns stdout/stderr

The Code Interpreter runs Python in an isolated sandbox directory with a 30-second execution timeout. It supports:

• Standard output capture
• Matplotlib chart generation (with Agg backend, no display required)
• Automatic print() wrapping for bare expressions

The Python executable path is configured in Administration → Settings → Python Path.

Prompt examples:

• "Calculate the compound interest on $10,000 at 5% for 10 years" → triggers execute_python
• "Generate a bar chart showing monthly sales: Jan=100, Feb=150, Mar=120" → triggers execute_python with matplotlib

Execution modes: NATIVE vs DOCKER

You choose how Python actually runs in Console → Global Settings → Code Interpreter:

Mode	What it is	When to use
NATIVE (default)	A host subprocess running the configured Python	Trusted content, simplest setup
DOCKER	Each execution runs in a fresh, hardened container	Untrusted input, multi-tenant, defense-in-depth

DOCKER is the recommended posture when code might be influenced by untrusted input. Every execution gets a throwaway container with baseline hardening always applied (--cap-drop ALL, --no-new-privileges, --read-only, and Docker's built-in seccomp profile) plus configurable limits under turing.code-interpreter.docker.*:

Key	Default	Purpose
network	none	none = no egress; or a named docker network
memory	512m	hard RAM cap per execution
cpus	1.0	CPU quota per execution
pids-limit	128	max PIDs (fork-bomb containment)
runtime	(host default)	set runsc for gVisor extra isolation (opt-in)
seccomp-profile	(Docker default)	path to a tighter seccomp profile (opt-in)

A "Check Docker" probe in Global Settings verifies the daemon is reachable before you switch.

Resource limits for the NATIVE path

DOCKER caps via the container runtime; the NATIVE path can stop a runaway script (e.g. [x*x for x in range(10**9)]) from exhausting host RAM before the timeout fires. Opt in under turing.code-interpreter.native.limits (Linux only):

turing:
  code-interpreter:
    native:
      limits:
        enabled: true       # default false (timeout is the only guard)
        memory-max: 1g       # hard RAM cap
        cpu-seconds: 35      # CPU-time cap
        limiter: auto        # auto | prlimit | systemd-run | none

Warm pool (NATIVE only)

A frequently-invoked tool pays a ~200ms cold-boot each call. Enable a pre-warmed pool of single-use interpreters to skip it:

turing:
  code-interpreter:
    warm-pool:
      enabled: true   # default false
      size: 2          # interpreters kept booted and ready

Workers are single-use (one execution then exit, no cross-tenant state leak) and refilled asynchronously. The pool is automatically bypassed in DOCKER mode and whenever the NATIVE limits above are enabled (a pooled worker can't be wrapped by prlimit/systemd-run).

Structured output & sandbox: URLs

Beyond stdout/stderr, the interpreter can return a structured result, used by Custom Tools via the Groovy code.executePythonStructured(...) / code.executePythonJson(...) bindings. It carries stdout, stderr, an exitCode, timedOut, durationMs, a rendered markdown, and a list of generated files (name, url, image, sizeBytes). Files (a chart PNG, a generated PDF) are addressed with Turing's sandbox: URL scheme (the OpenAI convention): clients resolve them to a real download URL, and the underlying /api/v2/code-interpreter/{session}/{file} URLs are HMAC-signed and cookie-bound so they can't be replayed across browsers. This retires the old "regex-parse the markdown for file links" approach.

Agents can also declare extra Python dependencies (e.g. reportlab matplotlib qrcode) that Turing ensures are installed before running.

Tool output feeding a slot

A Code Interpreter file can land in a multi-modal slot. Slot types IMAGE / AUDIO / FILE hold an artifact's storage URL, so a generated chart or document becomes a first-class value in a Chat Flow.

---

External Tools via MCP Servers

Beyond the native tools, AI Agents can access tools from any external server implementing the Model Context Protocol (MCP). This covers company-internal systems, proprietary data APIs, and the growing ecosystem of public MCP servers.

See MCP Servers for configuration details.

---

Provider-native server tools

The tools above are Turing's own, Turing runs them and feeds results back to the model. Separately, some vendors expose server-side built-in tools that run inside the provider's infrastructure (the model searches the web, executes code, or fetches a URL without Turing doing the work). These are selected through the capability gate, not the agent's tool list, and they coexist with Turing/MCP tools in one tool loop.

Function	OpenAI	Anthropic	Gemini
Web search	✅	✅	✅ (Google Search grounding)
Web fetch / URL context	n/a	✅	✅
Code execution	✅	✅	✅
Image generation	✅	n/a	✅
Computer use	✅	✅	✅
Remote MCP	✅	✅	n/a

Gemini's native primitives (Google Search grounding, URL Context, code execution, image generation, Computer Use, plus thinking budget, context caching, Batch, Files, full-context answering, and native video understanding) are documented in depth in Generative AI → Gemini native primitives. For how server-native and Turing tools share one loop, and the two-level gate that controls them, see Capabilities.

Long and reasoning-heavy tool loops

Two cross-vendor request options shape the tool loop itself:

• Interleaved thinking (Anthropic) lets Claude reason between tool calls, not just before the first one: useful for multi-step agentic loops.
• Background mode (OpenAI Responses) runs a long tool loop asynchronously and resumes by polling, so a deep code-interpreter or computer-use loop doesn't hold the HTTP connection open.

Both are opt-in and fail-open. See AI Agents → Reasoning, caching & background execution.

---

Large results are offloaded automatically

Some tools return a lot: a big search dump, a long catalog, a wide JSON array. Pasting all of that into the prompt every turn is wasteful. When a tool result exceeds turing.genai.tool-result-offload.inline-max-chars (default 4096), Turing writes the full payload to the Agent Workspace and replaces the inline result with a short workspace://… reference. The model pulls the data back only when it needs it, via an always-present workspace_read tool.

This applies uniformly to native, MCP, and Custom tools, and is active only when a storage backend is configured, so the default deployment behaves exactly as before. See Agent Workspace → Auto-offload for details.

---

Related Pages

Page	Description
AI Agents	How to compose agents with the tools they need
Custom Tools	Author Groovy tools; the code.executePython* bindings
Chat Flow	functionCall nodes and multi-modal slots fed by tool output
MCP Servers	Extend agents with external tools via MCP
Agent Workspace	Where large tool results are offloaded
DSL Query API	Full reference for the Elasticsearch-compatible DSL
DSL Compatibility Matrix	Engine compatibility for all DSL features
Assets	Knowledge Base files queried by RAG tools
Semantic Navigation	The search experience powering DSL tools

---