get_xbrl_facts

Raw XBRL facts from SEC filings — use only when get_financials cannot answer the question.

**Scope:** escape-hatch for dimensional / industry-specific / as-filed numbers. ~3,000 facts per filing with dimensional breakdowns (segment, geography, product line). Search by human-readable label, not XBRL concept names.

**First try get_financials** — 323+ standard metrics across TTM/FY/Q + YOY/CAGR for all 5,500+ companies: faster, cheaper, more portable across tickers.

**Use get_xbrl_facts only when:**

• You need a segment / geographic / product-line breakdown that get_financials aggregates away — available only for concepts the filer XBRL-tags dimensionally (usually revenue + segment profit / Adjusted EBITDA). Segment **costs** are frequently NOT tagged; a relaxed-search miss returns the recipe for recovering them.
• You need **revenue concentration / share** by customer, channel, distributor, geography, or product — the as-filed ConcentrationRiskPercentage facts. Deterministic, and present even when the prose only describes the relationship qualitatively. Search concentration.
• You need an industry-specific metric not in the standard catalog (e.g., medical cost ratio for a health insurer, reserve replacement ratio for an oil & gas name)
• You need to verify a specific number from filing text against the as-filed XBRL value
• You need a historical fiscal year not returned by get_financials (pass fiscal_year)
• You need a cash-flow / income **line across periods to de-cumulate a standalone quarter** — the as-filed cash-flow statement is cumulative YTD. Set period_history: true (see that parameter for the subtraction recipe).

**Sign — read this before quoting a direction.** <raw_data>.value is the raw as-filed XBRL instance value. It is a positive magnitude for outflow / contra-asset concepts (PaymentsTo…, capex, accumulated depreciation) but is **signed by construction** for the IncreaseDecreaseIn… working-capital family. The table shows that value, except where the instance value is negative AND carries the negated-label role — there it shows the filed face instead (so a working-capital *decrease* reads as the positive source of cash the statement shows).

(filed −) is a **best-effort hint, not a guarantee**: it comes from the filing's label linkbase and is recorded PER CONCEPT, so it fires on every use of a concept that is presented negatively *anywhere* in that filing — including lines the filing presents positively. Real example: Texas Instruments tags IncomeTaxExpenseBenefit with a negated label in its AOCI reclassification table, so its income-statement "Provision for income taxes" carries (filed −) even though the statement face reads 709, not (709).

**So: trust the value, treat the flag as a prompt to check.** When direction is load-bearing, confirm against the edgar fact link — or use get_financials, which applies a curated statement-sign map instead of the linkbase flag.

**Exactness:** the markdown rounds for readability; the <raw_data> block carries each fact's EXACT value + decimals (the as-filed precision floor: −3 = thousands, −6 = millions) + edgar — a deep-link to that exact fact in the SEC EDGAR iXBRL document, so every value resolves to the independent primary source.

**Share counts (pick the right basis):** the **dei** EntityCommonStockSharesOutstanding (shown as "cover-page / current") is the most-current shares outstanding, as of the filing/cover date — use it for **market cap, equity value, and "shares outstanding from the cover page"**. The **us-gaap** CommonStockSharesOutstanding is the **balance-sheet period-end** count (an earlier date); WeightedAverageNumberOf…SharesOutstanding is the **per-period average for EPS**. These can differ a few % for buyback-heavy / recently-issuing filers — match the as-of date in the Period column to your task.

Searching: commas OR-match (revenue,product); SPACES WITHIN a term AND-match, across the concept name, its label, and the filer's own dimension axis/member labels. A multi-word search is therefore a filter on the FILER's wording, not on meaning — it can drop the exact series you want with no warning. Start with ONE word and narrow client-side. Responses cap at ~20K chars — over it, whole facts are paged out of BOTH the table and the <raw_data> block together (the JSON stays valid) with an explicit facts_omitted count; narrow the search for the rest.

**No vantage_date** — the DEFAULT path resolves the latest filing as of TODAY. A past filing IS exactly reachable: pin it with accession_number (from list_filings, which takes vantage_date) or fiscal_year/fiscal_period — pin the filing rather than assuming latest for as-of work.