What Data for This Category Looks Like
The data for solid waste treatment intelligent due diligence reports mainly comes from enterprise hazardous waste management systems, ecological environment department supervision ledgers, third-party test reports, and transfer and transportation documents. Data updates follow individual project cycles, and are synchronized with project progress or quarterly compliance inspections. Most documents are structured multi-page files, containing fields such as hazardous waste category, production volume, disposal method, compliance qualification number, and disposal site coordinates. Units include tons, cubic meters, milligrams per cubic meter and other professional measurement standards. Some fields must match the unified coding rules of ecological environment departments.
Constraints Imposed by These Characteristics on Deployment and Upgrade
Solid waste treatment due diligence data includes multiple types of professional measurement fields and compliance codes. During deployment, adapt to the large-field storage and coding parsing rules of the vector database to avoid field parsing failures. Data update cycles are flexibly adjusted according to project progress. During upgrades, support incremental synchronization configuration to avoid excessive server resource occupation from full synchronization. Some fields must match the latest coding rules of ecological environment departments. During deployment, preset field verification logic. During upgrades, synchronously update the mapping rules for compliance verification. Parsing of multi-page structured documents must support custom field extraction. During deployment, reserve a configuration entry. During upgrades, quickly adapt to new regulatory document formats.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 300-600 seconds | Solid waste treatment due diligence reports are mostly multi-page long documents, with long parsing times. This range avoids timeout interruptions of parsing tasks |
maxContext | 8000-12000 token | The documents contain multiple sections of professional content and compliance fields, requiring sufficient context to accommodate complete parsed information |
Recall Count | Top 8-12 entries | Compliance fields in due diligence reports are scattered across different document chunks. A sufficient recall volume is needed to cover key information |
Similarity Threshold | 0.75-0.85 | Solid waste treatment data contains many professional terms. This range balances recall precision and coverage, avoiding omission of key compliance-related content |
UPLOAD_FILE_MAX_SIZE | 2000 MB | A single solid waste treatment due diligence report may include multiple attachments such as test reports and transportation documents. This setting supports large file uploads |
rerank_top_n | Top 5-8 entries | Key compliance information in due diligence reports is concentrated. Reranking filters irrelevant content and improves answer accuracy |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Configuration Mistakes
- Symptom: When using the Qwen2.5 model deployed via Ollama, tool call triggers fail, and the interface returns an error stating the model does not support tool calls. Cause: The tool call switch is not enabled in the model configuration, or the Ollama-deployed model has not loaded the tool call plugin.
- Symptom: Only the FastGPT Docker container image is upgraded, without updating the supporting model proxy configuration. This causes model call interface errors. Cause: The model proxy layer and FastGPT's interface call protocol have a version binding relationship. Upgrading the main program alone will lead to protocol incompatibility.
- Symptom: After setting the knowledge base chunk size to 5000 tokens and the retrieval reference upper limit to 1500 tokens, retrieval results still include document chunks exceeding 1500 tokens. Cause: The reference upper limit only constrains the total length of the final spliced context, and does not limit the return length of a single recalled chunk. The chunk truncation rule was not configured synchronously.
How to Verify Proper Configuration
- Upload a sample solid waste treatment due diligence report, and confirm that the fields extracted after parsing match the preset business field list.
- Initiate a retrieval query targeting compliance fields, and confirm that the number of returned document chunks matches the configured recall count.
- Verify that models deployed via Ollama can normally trigger tool calls, with no call failure errors.
- Check the Docker container logs, confirm that there are no continuous timeout errors in parsing tasks, and that the service runs stably.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.