daymade/claude-code-skills/bigdata-skill

Bigdata.com SDK + REST Toolkit

Get the structured substrate the Bigdata.com MCP server doesn't hand over. The MCP returns clean prose and pre-synthesized tearsheets, but its search tool gives chunks with no per-chunk sentiment or entity spans, and its tearsheets give aggregate values — not the fiscal-period time series, universe screener, or per-field JSON you'd build a pipeline on. The official bigdata-client SDK plus a thin REST passthrough over the same backend, same JWT reach the official /v1/* endpoints that hold it. This skill bundles a toolkit that does exactly that — already debugged, already cost-guarded — so you don't re-pay the discovery cost.

The core problem this solves (read this first)

The Bigdata MCP server answers "what's the sentiment around NVIDIA?" with a readable paragraph or a pre-synthesized tearsheet — genuinely useful for a chat turn. But the moment you need the machine-readable substrate to build a pipeline on, the MCP doesn't hand it over:

• its search tool returns chunks with text + relevance only — **no per-chunk

sentiment number, no entity character spans**;

• its tearsheets give aggregate values (a single sentiment score, a summary

of estimates) — not a fiscal-period time series you can compute on, a universe screener, or per-field JSON.

The fix is a general pattern, not a Bigdata trick:

**When an MCP data source returns only synthesized output but you need the

structured fields underneath, drop to the vendor SDK or REST.** MCP optimizes

for a chat turn, not a pipeline.

Crucially, for Bigdata these structured fields are official, publicly documented REST endpoints (docs.bigdata.com/api-reference/...), not a hidden backend — and Bigdata is sunsetting the SDK (EOL 2026-12-31) in favour of this REST API, so the REST layer here is the forward-compatible path, not a hack. The SDK (bigdata_client.Bigdata) covers search + knowledge-graph; `bd._api.http` reaches every /v1/* endpoint the SDK never wrapped. The bundled bigdata_toolkit packages both behind one BigdataClient.

When to use this skill

Trigger on any of these, in any language:

• The user is using Bigdata.com / RavenPack and the MCP result feels thin —

"where's the sentiment score?", "I need entity-level data", "the calendar".

• They want forward / structured financials for a ticker: analyst

estimates, earnings or event calendar, earnings surprise, analyst ratings, price targets, a company screener / universe.

• They want annotated news chunks with numeric sentiment + entity spans, or

a sentiment time series / co-mention graph.

• They mention a `bd_v2_` API key, rp_entity_id, query_unit / chunk

cost, bigdata-client, or "the bigdata MCP isn't enough".

• They're building an investment-research dataset and need a reusable,

cost-aware data-pull layer rather than one-off MCP calls.

Setup (one time)

1 — API key (never hardcode it). The client fail-fasts if it's missing:

export BIGDATA_API_KEY=bd_v2_xxxxxxxx

2 — An isolated Python env with the official SDK. The bundled toolkit imports bigdata_client; install it once:

uv venv .venv --python 3.12
uv pip install --python .venv/bin/python bigdata-client
# Behind a slow/blocked PyPI (e.g. mainland China) add a mirror, and unset any
# outbound proxy for the install step so uv reaches the index directly:
#   --index-url https://pypi.tuna.tsinghua.edu.cn/simple

3 — Outbound proxy (only if your network needs one to reach `api.bigdata.com`). Two equivalent options — the official SDK accepts both: an env var, or BigdataClient(proxy=...) in code. The env var is simplest:

export HTTPS_PROXY=http://<host>:<port>     # plus WSS_PROXY for chat/WebSocket

If a proxy does TLS interception (self-signed CA) and you hit SSL handshake errors, the official fix is BigdataClient(verify_ssl="<proxy-CA>.pem") — not blind retries.

4 — Make the bundled package importable by putting this skill's scripts/ on PYTHONPATH (or sys.path.insert(0, "<this-skill>/scripts")).

Smoke-test the whole path (entity resolve + quota are free; --with-search adds one ~1 query_unit chunk search):

BIGDATA_API_KEY=bd_v2_xxx PYTHONPATH=scripts .venv/bin/python scripts/probe_example.py

Quickstart

import sys
sys.path.insert(0, "<this-skill>/scripts")          # so `import bigdata_toolkit` resolves
from bigdata_toolkit import (
    BigdataClient, EntityResolver, AnnotatedSearcher,
    StructuredDataREST, CostTracker, CostModel, rc,   # rc = SSL-retry wrapper
)
c  = BigdataClient()                                  # SDK + REST escape hatch, one object
er = EntityResolver(c)
nvda = rc(lambda: er.resolve_id("NVIDIA", country="US"))   # -> 'E09E2B'  (rp_entity_id is the gateway key)
# --- Structured financials the MCP does NOT expose (REST escape hatch) ---
rest = StructuredDataREST(c)
est  = rc(lambda: rest.analyst_estimates(nvda, period="quarter", limit=5))  # forward consensus
surp = rc(lambda: rest.latest_surprise(nvda))                               # last EPS/revenue surprise
cal  = rc(lambda: rest.events_calendar(nvda, categories=["earnings-call"],
                                       start_date="2026-06-01", end_date="2026-12-31"))
# --- Annotated chunks the MCP STRIPS: sentiment + entity spans (cost-guarded) ---
s    = AnnotatedSearcher(c)
docs = rc(lambda: s.search_entity(nvda, keyword="data center", chunk_limit=10))
# each chunk dict: {"sentiment": float, "entities": [{"key": rp_id, "start", "end"}], "text", ...}
# --- Always know your spend (chunk-billed; see Cost discipline) ---
ct = CostTracker(c); ct.snapshot()
# ... run a batch ...
print(ct.delta())     # {'delta_chunks':..., 'delta_query_units':..., 'usd_fast':...}

Wrap every network call in rc(lambda: ...) — a first-handshake SSL: UNEXPECTED_EOF is common and the SDK's internal retry doesn't cover it.

Routing — which capability answers the question

income/balance/cash-flow/daily-prices/dividends/revenue-segments return

{fields, values} — wrap them in fields_values_to_records() to get

[{field: value}]. The *_ttm / company_profile endpoints are already flat.

All structured endpoints above are free (0 chunks) except

connected_entities and AnnotatedSearcher (chunk-billed).

The two data faces (do NOT say "Bigdata fails for Chinese / A-shares")

This split is the most important non-obvious conclusion — state it precisely:

Cost discipline

1 query_unit = 10 chunks (official). Only chunk-search is billed — the structured /v1/* endpoints (estimates, financials, prices, calendar, surprise, ratings, the sentiment time series, screener…) are free (0 chunks, contract-tested). connected_entities (co-mentions) and AnnotatedSearcher are chunk-billed.

Three levers when you do pay for chunks:

1. `ChunkLimit`, never a bare `int`. Search.run(int) is a document limit

billed by the full chunk page; ChunkLimit(n) bills per chunk. AnnotatedSearcher.search forces ChunkLimit for you. (We observed roughly a 52x gap once — a single measured data point, not stated in the official docs; treat the exact multiple as indicative. The rule "use ChunkLimit" holds regardless, because max_chunks is the official billing unit.)

1. Rerank bills only the *returned* chunks (official) — pass a

rerank_threshold to recall broadly but pay only for the high-relevance hits.

1. Batch search is 50% cheaper ($0.0075 vs $0.015 / qu) — use

BatchSearch for a large multi-query backfill.

Use CostModel to veto an over-budget job before running it, and CostTracker.snapshot() / delta() to measure real spend. Full accounting → references/cost_accounting.md.

Known pitfalls (already solved — don't re-debug these)

Each cost real debugging time and is fixed or guarded in the toolkit. Full reproductions and fixes in `references/known_pitfalls.md`:

1. First-handshake `SSL: UNEXPECTED_EOF` → wrap calls in rc(); the SDK's

urllib3 retry only covers HTTP status, not the SSL EOF.

1. `All(entity, Keyword(kw))` raises `TypeError` → combine with the &

operator (entity & Keyword(kw)); All takes a single iterable. (Fixed in AnnotatedSearcher.entity_query.)

1. The 52x doc-limit billing trap → always ChunkLimit, never a bare int.
2. Closure capture in loops → bind loop vars: rc(lambda q=q, dr=dr: ...).
3. `analyst_estimates(period="quarter")` 400s above `limit≈20`.
4. `company_screener` filters must nest under `"filters"` — flat top-level

keys don't 400, they're silently dropped → unfiltered universe.

1. `Document.reporting_period` is always `None` (the SDK model drops a field

present on the REST wire) → fetch_reporting_period_raw.

What this skill will not do