Spaces:
Sleeping
Sleeping
File size: 4,740 Bytes
b2931f4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 | """Agent tools — provider-neutral registry.
Each tool is a plain Python function (independently testable) plus a `ToolSpec`
describing its name, when-to-use text, and JSON-schema parameters. The spec is
deliberately NOT in any vendor's tool format: Decision 16 adapts these into
LangChain/LangGraph tools (which bind to Gemini or Claude alike), keeping the
provider seam from Decision 14 intact. Hardwiring Anthropic's tool_use shape
here — as the original handoff assumed — would have undone that.
`dispatch(name, args)` is the single call site the agent loop uses to run a
tool by name; it returns the tool's dict result unchanged.
Three tools, by design — see [[project_finrag_overview]]:
calculator — arithmetic, safe-eval (pure fn)
lookup_citation — re-fetch a chunk from Qdrant (read-only)
sql_query — NL → SQL over DuckDB facts (sub-LLM; added next)
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Callable
from finrag.tools.calculator import calculator
from finrag.tools.citation import lookup_citation
from finrag.tools.sql import sql_query
@dataclass(frozen=True)
class ToolSpec:
name: str
description: str
# JSON-schema "object" describing the args. Both Gemini and Anthropic
# accept this shape (modulo a thin adapter), as does LangChain.
parameters: dict[str, Any]
fn: Callable[..., dict[str, Any]]
TOOL_SPECS: list[ToolSpec] = [
ToolSpec(
name="calculator",
description=(
"Evaluate an arithmetic expression over numeric literals "
"(+ - * / // % ** and parentheses). Use for growth rates, margins, "
"sums, and ratios instead of doing mental math. Extract the numbers "
"from context first, then pass an expression like "
"'(383285 - 394328) / 394328 * 100'."
),
parameters={
"type": "object",
"properties": {
"expression": {
"type": "string",
"description": "Arithmetic expression over numeric literals only.",
}
},
"required": ["expression"],
},
fn=calculator,
),
ToolSpec(
name="lookup_citation",
description=(
"Re-fetch the full text and provenance of a single retrieved chunk "
"by its chunk_id. Use when you need to quote an exact figure or "
"re-read a chunk you cited earlier."
),
parameters={
"type": "object",
"properties": {
"chunk_id": {
"type": "string",
"description": (
"The chunk_id of a previously retrieved chunk — the exact "
"value shown as (id=...) in that chunk's header. Do not "
"use the [N] anchor or invent an id."
),
}
},
"required": ["chunk_id"],
},
fn=lookup_citation,
),
ToolSpec(
name="sql_query",
description=(
"Query exact financial figures (revenue, net income, R&D, margins, "
"multi-year or cross-company comparisons) from the structured "
"financial_facts database. Pass a natural-language description of "
"the numbers you need; SQL is generated and run for you. Prefer this "
"over reading figures out of text chunks when precision matters."
),
parameters={
"type": "object",
"properties": {
"natural_language": {
"type": "string",
"description": "Plain-language description of the figures to fetch.",
}
},
"required": ["natural_language"],
},
fn=sql_query,
),
]
# name → spec, for O(1) dispatch.
TOOL_REGISTRY: dict[str, ToolSpec] = {spec.name: spec for spec in TOOL_SPECS}
def dispatch(name: str, args: dict[str, Any]) -> dict[str, Any]:
"""Run tool `name` with keyword `args`. Returns the tool's dict result.
Unknown tool names return an error dict (not a raise) so a hallucinated
tool call degrades gracefully inside the agent loop.
"""
spec = TOOL_REGISTRY.get(name)
if spec is None:
return {"error": f"Unknown tool {name!r}. Available: {list(TOOL_REGISTRY)}"}
try:
return spec.fn(**args)
except TypeError as e:
# Wrong/missing args from the model — surface as data, not a crash.
return {"error": f"Bad arguments for {name!r}: {e}"}
__all__ = ["ToolSpec", "TOOL_SPECS", "TOOL_REGISTRY", "dispatch"]
|