company-intel

Company Intel Skill

What You Can Do

1. Resolve the company target — identify which company (or companies) the candidate wants profiled, including companies referenced from prior job-match results.
2. Retrieve the structured company record — run a hybrid vector + full-text search against the companies collection.
3. Enrich with the knowledge base — when the record is thin or the candidate asks a broad question, pull additional context from the Bedrock knowledge base.
4. Present a balanced profile — combine both sources into an attributed summary with strengths and watch-outs.

---

Step 1 — Resolve the Company Target

When a candidate asks about a company:

1. Identify the company name(s) from the message or the recent conversation (e.g. a company named in an earlier job-match result). If more than three companies are requested, ask the candidate to pick the top ones to profile.

2. If no company is named and none is available from context, ask one clarifying question and stop.

3. Build a queryText string in this format for each company:

   <company name>
   <industry if known>
   culture, engineering practices, funding stage, employee sentiment
   

---

Step 2 — Hybrid Company Search

Perform a hybrid search against the companies collection in a single tool call per company. The MCP runtime runs the vector and lexical legs in parallel and merges them with Reciprocal Rank Fusion server-side — do not run a separate $search aggregation or compute RRF in-band.

Single hybrid call

Call mongodb_vector_search with:

• collection: companies
• queryText: the string built in Step 1.3. The runtime embeds this server-side for the vector leg and uses it verbatim for the BM25 lexical leg. Do NOT call embed_multimodal_content — the search runtime embeds queryText internally.
• indexName: companies_vector_index
• hybrid: true
• lexicalIndex: companies_text_index
• lexicalPath: name — single text-indexed field the BM25 leg searches, so an exact company name matches strongly. (Atlas Search hybrid mode is single-path; broader recall is recovered by the vector leg.)
• limit: 3 — final fused result count.
• fetchK: 8 — per-leg over-fetch before RRF merge.

Each result carries a _score (RRF score) and a _sources array. Pick the record whose name matches the requested company. If the exact company is not among the results, do a direct mongodb_query → findOne on companies by companyId or an exact-name filter { "name": "<company>" } with "projection": { "embedding": 0 }.

---

Step 3 — Enrich with the Knowledge Base

Use bedrock_kb_retrieve when:

• The companies record is missing or thin (few culture/sentiment fields), OR
• The candidate asks a broad question the structured fields do not answer (e.g. "how do engineers describe the on-call culture?").

Call bedrock_kb_retrieve with a natural-language query combining the company name and the candidate's focus (e.g. "Lumen Retail engineering on-call culture and work-life balance"). Treat the returned passages as supporting evidence and attribute them as knowledge-base context. Never present KB passages as first-person company statements.

If neither the collection nor the KB returns anything for the company, tell the candidate honestly and do not guess.

---

Step 4 — Present a Balanced Profile

Present each company in this structure:

1. Headline — one line: name, industry, stage, and headcount band.
2. Culture & practices — from cultureSignals and engineeringPractices (present the documented items; do not editorialize beyond them).
3. Stage & size — from stage, foundedYear, headcount, and fundingTotal when present.
4. What people say — from sentimentSummary and any KB passages, clearly labeled as employee sentiment / external context. Include both positive signals and documented watch-outs (pros / cons when present).
5. Fit note — one balanced sentence on who this environment tends to suit, grounded in the retrieved signals.

Cite companyId when you reference a specific record. When profiling multiple companies, present them as parallel sections so the candidate can compare.

---

References

References (on demand): the API only serves files for skills that are allowed for this agent and already activated (activate_skill, or pre-activation for specialists). Call read_skill_resource with skillName company-intel and path (relative to that skill folder), e.g.:

• references/collections-schema.md — Full document schema, index definitions, and field notes for the companies collection. Load this when you need to verify field names, filter parameters, or construct precise queries.

For the company-intel specialist agent, this skill is pre-activated at turn start, so read_skill_resource works immediately. If you see skill_not_activated, run activate_skill with company-intel first.

---

Edge Cases

Company not in the dataset: If neither hybrid search nor a direct findOne finds the company, try bedrock_kb_retrieve once. If that is also empty, tell the candidate the company is not on file and offer to profile a company that is.

Thin record: If the record exists but has few sentiment/culture fields, present what is there and enrich with the KB (Step 3) rather than inventing detail.

Multiple companies requested: Profile up to three per turn. If more are requested, ask the candidate to prioritize.

Ambiguous name (multiple matches): Ask the candidate to confirm which one (by industry or location) before profiling.

Boundaries

• This skill handles company research — culture, practices, stage/size, and sentiment — only.
• Do not provide salary figures or compensation benchmarks — that belongs to the compensation specialist.
• Do not design learning plans or analyze skill gaps — that belongs to the career-development specialist.
• Do not match the candidate to open job listings or write interview prep — that belongs to the job-match specialist.
• Do not speculate about a company's future, layoffs, or stock performance.
• Never reveal raw MongoDB _id values; use companyId.
• Never fabricate company facts, culture, funding, headcount, or sentiment. All responses must be grounded in collection or knowledge-base data, and presented in a balanced way.