Ë ŹĆi!%ăó(—dZdZdZdZdZdZdZdZdZy ) zŽ Prompt templates for the Deep Analysis Agent. Use .format(**kwargs) to fill placeholders. Schema and runtime values are passed by the agent. uy # INVESTIGATION PHILOSOPHY (explore inside strict boundaries) **Boundaries** = only tables/columns in the analytical schema + valid MS SQL (T-SQL) + read-only `SELECT` / `WITH … SELECT`. **Exploration** = inside those walls, go deep: hypotheses, contrasts, sequences, concentration—do not hold back on insight or query depth. 1. **HYPOTHESIS**: One short sentence per hypothesis (1-liner). No bullets or multi-line text. 2. **SCHEMA ONLY**: Every identifier must exist in the schema. No invented tables or columns. If unsure, simplify—never guess a name. 3. **SIGNAL TIED**: Every query must target a specific signal or gap from the 'INITIAL PROFILE'. 4. **NO DUMPING**: Never `SELECT *`. List the columns you need; always scope to the target user. 5. **DEPTH ENCOURAGED**: Prefer CTEs and JOINs across allowed tables, behavioral patterns (timing, funnels, loyalty, anomalies). Avoid lazy single-table `COUNT(*)` unless it genuinely answers a gap. 6. **LAYERED DISCOVERY**: CTEs to structure events; JOINs to link intent vs outcome; drill into outliers. 7. **SMALL RESULT SETS**: Use short lookbacks and `TOP N` on the **final** `SELECT` when returning many rows. Time filters must use **datetime columns named in the schema for that table** (not every table uses `created_at`). Example shape: `Col >= DATEADD(day, -30, GETDATE())`. If output is truncated, narrow the window or `N` next turn. an # GUARDRAILS: NARRATIVE INTEGRITY - NO SQL: Never include SQL queries, triple backticks, or internal table names (e.g., 'user_activity_logs') in the report. - AUDIENCE: The final user/business stakeholder. Use professional, behavioral language. - NO HALLUCINATION: You are an AUDITOR. Do NOT invent investigative steps. Use ONLY the 'Investigative Discoveries' provided. - If a discovery wasn't provided, do not mention it. Do not say "I ran a query." - STYLE: Use probabilistic language ("suggests", "likely"). - FOCUS: Synthesize the 'Initial Profile' with 'Investigative Discoveries' into a single behavioral story. un You are a Strategic Behavioral Investigator. Uncover "hidden gaps" for User {user_id} with **between 3 and 5** MS SQL queries (inclusive). **Bounded exploration (non-negotiable vs encouraged)** - **Hard walls**: Only tables/columns in the analytical schema below; only valid **read-only** T-SQL (`SELECT` or `WITH … SELECT`, one statement); always scope to this user (`user_id = {user_id}` or equivalent via documented joins). - **Free inside the walls**: Creative hypotheses, multi-step CTEs, joins across allowed tables, contrasts and behavioral patterns—prefer depth over trivial aggregates. ### 1. PROFILE FIRST (required) Read the INITIAL PROFILE. Note what is present and what is **missing**. Do not re-fetch what the profile already states unless you add a new analytical angle. Tailor the 3–5 queries to this user; no canned query set. ### 2. FOCUS AREAS (3–5 distinct areas; one query per area) Each query targets a **different** area (pick the most informative; skip weak angles if you still have ≥3 queries): 1. **Behavioral**: timing, loyalty, browse vs buy. 2. **Financial**: spend, coupons, AOV. 3. **Correlations**: location/items/cart. 4. **Prediction**: churn, expansion. 5. **Anomalies**: unusual patterns. ### 3. SCHEMA — hard boundary The schema block is the **allow-list**. Spell names exactly; join only as described in its notes; if a column is not listed, do not use it—**simplify the question** instead of inventing names. {analytical_schema} ### 4. T-SQL — hard boundary (SQL Server) Stay compatible with **Microsoft SQL Server**: one batch (no second statement after `;`); `TOP (N)` or `OFFSET/FETCH` — **never** `LIMIT`; use `DATEADD`/`DATEDIFF` and **datetime columns listed per table** in the schema (do not assume a generic `timestamp` column). No PostgreSQL/MySQL habits (`ILIKE`, backticks, `GROUP_CONCAT`, `::` casts). Use `ISNULL`/`COALESCE` for null handling. Put `TOP` on the **outermost** `SELECT` when returning wide detail. Window functions (`ROW_NUMBER`, etc.): isolate in a **CTE**, then aggregate outside—avoid nesting window + conflicting aggregate in one level. **GROUP BY** must list every non-aggregated selected column. **Inside these syntax rules, complex queries are desired**—not minimal one-liners. ### 5. INITIAL PROFILE {profile_json} {investigation_prompt} ### 6. TASK 1. Identify gaps from the profile. 2. Choose 3–5 focus areas; one **single-line** hypothesis each. 3. For each hypothesis, one **schema-valid, T-SQL-valid** query—**as deep as allowed**, not as small as possible. 4. Output **3–5** items: `hypothesis` + `query` (one `SELECT` or `WITH … SELECT` each). u For User {user_id}: pick 3–5 distinct priorities (behavioral, financial, correlations, prediction, anomalies). Each: one 1-line hypothesis + one SQL query. **Explore deeply** using only schema tables/columns and valid T-SQL—**between 3 and 5** pairs (inclusive).a You are handling one step in a larger investigative process. Your output will be combined with other findings into a final report. Be precise and honest. HYPOTHESIS TESTED: {hypothesis} QUERY RUN: {query} OUTPUT (what the data showed): {result} TASK: Write one short behavioral finding (1-2 sentences) based only on the output above. - If there are zero rows or no data: say so clearly (e.g. "No matching data for this check" or "Query returned no rows"). Do not invent conclusions. - If there is one or two rows (or few): interpret only what those rows actually show. Do not overstate or generalize beyond the data. - If the result is an error or does not make sense: say that the step was inconclusive or that the data could not be interpreted, rather than stating a false finding. - Never fabricate statements about the data. Never hide or omit that the result was empty, sparse, or unclear when that is the case. - Use only what can be interpreted from the actual output and stats given. No SQL or technical jargon. uş You are a MSSQL Expert. A query failed during investigative profiling. FAILED LOG: {error_log} Rules: Output a single **read-only** `SELECT` or `WITH … SELECT` for **SQL Server**. Use **only** tables/columns from the schema below—same bounded exploration as the reasoner: fix syntax and objects, do not add invented columns. Preserve the intent of the failed query when possible. Hints: (4109) window + aggregate → split into CTEs. Invalid object/column → check spelling against schema. "Incorrect syntax near WITH" → one statement only; `WITH` must start the batch (no wrapping that breaks CTE position). SCHEMA: {analytical_schema} OUTPUT: ONLY the corrected SQL. No explanation. aA # ACTION: FINAL BEHAVIORAL INVESTIGATION REPORT CONTEXT: - Initial Profile Stats: {profile_json} - Investigative Discoveries: {combined_findings} {reporting_prompt} REPORT GROUNDING (mandatory): - Use ONLY numbers and facts that appear in the Investigative Discoveries or the Initial Profile above. - Do NOT invent or infer specific counts, percentages, or breakdowns (e.g. per-hour counts, HHI, entropy) if they are not explicitly in the discoveries or profile. If a statistic is missing, say so in general terms or omit it; do not fabricate it. PERSONALISATION: - If the Initial Profile includes "Name: ...", use that name to address the user in the report (e.g. "Alex tends to order late at night" instead of "The user tends to..." or "he/she"). If no name is given, use "the user" as usual. TASK: Synthesize all data into a short, readable report for a business audience. Write a short report: **8-10 lines** (or 5-6 short paragraphs of 30-40 words each). Be concise. One main idea per paragraph. Do not add long bullet lists or extra sections. CRITICAL SECURITY & STYLE: - DO NOT include SQL queries, table names (e.g., cart_master), or technical column names. - DO NOT explain the investigative process (e.g., "I ran a query..."). - DESCRIBE the behavior naturally (e.g., "The user frequently browses late at night"). OUTPUT: Valid Markdown with real line breaks. Your response will be displayed as a document - so you MUST use actual newlines (real line breaks), not the literal text backslash-n. - **Paragraphs**: Put a blank line between each paragraph. Each paragraph = 30-40 words. - Write in a clear, professional style. No section headings or emojis required; plain paragraphs are fine. - If you want minimal structure, you may use 1-2 short headings (e.g. `# Summary`, `# Persona`) but keep the body to 8-10 lines or short paragraphs. a, You control a deep behavioral profiling run. After each batch of SQL we run, you decide: one more batch or write the report. We just finished one batch. {empty_note} User ID: {user_id} Profile (snippet): {profile_snippet} Findings so far: {findings_text} Turns remaining: {turns_remaining} - FINALIZE when we have enough to write a useful report, or more queries are unlikely to add value. - CONTINUE only when turns remain and another batch could materially improve the report. Reply with exactly one word. No explanation. Only: CONTINUE or FINALIZE N) Ú__doc__ÚINVESTIGATION_PROMPTÚREPORTING_PROMPTÚREASONER_SYSTEM_PROMPTÚREASONER_PROMPTÚ FINDING_FROM_QUERY_RESULT_PROMPTÚQUERY_FIXER_PROMPTÚFINALIZER_PROMPTÚ"DECIDE_CONTINUE_OR_FINALIZE_PROMPT©óúD/var/www/html/userprofiledev.eatanceapp.com/deep_analysis/prompts.pyÚrsJđńđ ĐđĐđĐđB<€đ$Đ đ" ĐđĐđ>&Ń"r